Amazon

(AWS Certified AI Practitioner AIF-C01)

AWS Certified AI Practitioner AIF-C01

Total: 166 Questions

Question: 1
A company makes forecasts each quarter to decide how to optimize operations to meet expected demand. The
company uses ML models to make these forecasts.
An AI practitioner is writing a report about the trained ML models to provide transparency and explainability to
company stakeholders.
What should the AI practitioner include in the report to meet the transparency and explainability requirements?

A.Code for model training

B.Partial dependence plots (PDPs)
C.Sample data for training
D.Model convergence tables

Answer: B

Explanation:

The correct answer is B. Partial dependence plots (PDPs). Here's why:

Transparency and explainability in AI/ML refer to the ability to understand how a model makes its predictions.
This is crucial for trust, accountability, and debugging. While code, sample data, and model convergence
tables are valuable for model development and understanding its performance during training, they don't
directly address why the model makes a particular prediction for a given input.

Partial Dependence Plots (PDPs) directly contribute to explainability. A PDP visualizes the marginal effect of
one or two features on the predicted outcome of a machine learning model. It shows how the model's
prediction changes as the selected feature(s) vary, while holding all other features constant (in a sense,
averaging out their effects). By showing these relationships, stakeholders can gain insight into which features
are most influential in driving the model's forecasts and in what direction.

Code (A) is important for reproducibility, but it doesn't directly explain why the model is making certain
forecasts. Sample training data (C) is useful for verifying the data quality and distribution, but it doesn't reveal
the model's internal logic. Model convergence tables (D) show how well the model learned during training, but
don't illustrate feature importance or impact on predictions. Only PDPs effectively showcase the feature-
outcome relationships learned by the model, enhancing understanding for stakeholders. For example, a PDP
might show that as advertising spend increases, the predicted demand also increases, but the effect plateaus
at a certain point, giving actionable insights for decision-making.

Therefore, PDPs are the most relevant tool for meeting the transparency and explainability requirements in
the context of explaining ML model forecasts to company stakeholders. They enable stakeholders to
understand the model's reasoning and build trust in its predictions.

Relevant links:

Interpretable Machine Learning by Christoph Molnar: https://christophm.github.io/interpretable-ml-

book/pdp.html
scikit-learn PDP documentation: https://scikit-learn.org/stable/modules/partial_dependence.html

Question: 2
A law firm wants to build an AI application by using large language models (LLMs). The application will read legal
documents and extract key points from the documents.
Which solution meets these requirements?

A.Build an automatic named entity recognition system.

B.Create a recommendation engine.
 C.Develop a summarization chatbot.
D.Develop a multi-language translation system.

Answer: C

Explanation:

The correct answer is C. Develop a summarization chatbot.

Here's why: The law firm needs a solution that can process legal documents and extract key information. A
summarization chatbot is designed to ingest text and produce a condensed, coherent summary highlighting
the main points. LLMs are particularly effective for text summarization due to their ability to understand
context, identify crucial information, and generate human-quality summaries. A chatbot interface allows users
to interact with the LLM, specify documents, and receive the extracted key points in a conversational manner.

Option A, building an automatic named entity recognition (NER) system, while helpful for identifying entities
like names, organizations, and dates, doesn't inherently summarize or extract key points. It only identifies and
classifies pre-defined entity types. Option B, creating a recommendation engine, is irrelevant as it focuses on
suggesting items based on user preferences or historical data. Option D, developing a multi-language
translation system, addresses language translation and doesn't extract or summarize content.

Therefore, a summarization chatbot leverages LLMs to best satisfy the requirement of reading legal
documents and extracting key points, providing a user-friendly interface for accessing the summaries.

For more information on text summarization using LLMs:

Amazon SageMaker JumpStart text summarization: https://aws.amazon.com/sagemaker/jumpstart/ (search

for summarization)
Generative AI on AWS: https://aws.amazon.com/machine-learning/generative-ai/

Question: 3
A company wants to classify human genes into 20 categories based on gene characteristics. The company needs
an ML algorithm to document how the inner mechanism of the model affects the output.
Which ML algorithm meets these requirements?

A.Decision trees
B.Linear regression
C.Logistic regression
D.Neural networks

Answer: A

Explanation:

The correct answer is A, Decision Trees. Here's why:

Decision Trees are highly interpretable machine learning algorithms. Their structure mimics a flowchart, with
each internal node representing a test on an attribute (gene characteristic in this case), each branch
representing the outcome of the test, and each leaf node representing a class label (one of the 20 gene
categories). This clear, hierarchical structure allows users to easily trace the path from input features to the
final classification. You can see exactly which features were used to make a decision, and in what order.

The key requirement is understanding the model's inner workings. Linear regression and logistic regression,
while interpretable to some extent (coefficients show feature importance), don't provide the same level of
granular, step-by-step explanation as decision trees. Neural networks, especially deep neural networks, are
 notoriously "black boxes." It's difficult to understand exactly why a neural network made a specific prediction.

Decision Trees are also suitable for multi-class classification problems like classifying genes into 20
categories. While they might not always be the most accurate model (prone to overfitting), their
interpretability makes them ideal when understanding the why behind the prediction is crucial.

Here's a breakdown of why the other options are less suitable:

B. Linear Regression: Primarily for predicting continuous values, not suitable for multi-class classification.
While coefficients indicate feature importance, it doesn't detail decision paths.
C. Logistic Regression: Primarily for binary classification. While multi-class extensions exist, they still lack
the detailed, traceable paths of decision trees.
D. Neural Networks: Extremely difficult to interpret due to their complex, non-linear structure. Techniques
like LIME and SHAP can help with post-hoc interpretability, but don't inherently reveal the internal decision-
making process.

Decision tree's ease of visualization and inherent traceability of the path taken through feature evaluation
makes it a powerful tool when you need to understand and document the reasoning behind the ML model.

Authoritative Links:

Decision Trees: (Scikit-learn documentation - a popular Python ML library) - https://scikit-

learn.org/stable/modules/tree.html
Interpretability of Machine Learning: (Google AI Blog) - https://ai.googleblog.com/2023/03/interpretability-
beyond-feature.html (While this blog is a broader topic, it provides context on the value of interpretable
models.)

Question: 4
A company has built an image classification model to predict plant diseases from photos of plant leaves. The
company wants to evaluate how many images the model classified correctly.
Which evaluation metric should the company use to measure the model's performance?

A.R-squared score
B.Accuracy
C.Root mean squared error (RMSE)
D.Learning rate

Answer: B

Explanation:

Here's a detailed justification for why Accuracy is the most appropriate evaluation metric in this scenario:

The company needs to determine how often its image classification model correctly identifies plant diseases
in leaf photos. Image classification is a specific type of supervised machine learning problem where the model
predicts a category or class label for each input image.

Accuracy is the ratio of correctly classified instances to the total number of instances. In this case, it
represents the proportion of plant leaf images that the model correctly classified as having a specific disease
(or not). A higher accuracy score indicates better performance.

Here's why the other options are not suitable:

R-squared score: This metric is used to evaluate the performance of regression models, which predict
continuous values, not classification models.
 Root Mean Squared Error (RMSE): Similar to R-squared, RMSE is a regression metric, measuring the average
magnitude of errors between predicted and actual continuous values.
Learning rate: This is a hyperparameter that controls how much the model's weights are adjusted during
training. It doesn't measure the model's performance after training.

Accuracy provides a straightforward and easily interpretable measure of how well the model is performing at
its primary task: correctly classifying plant diseases based on images. This makes it the best choice for the
company's evaluation needs.

Therefore, the correct answer is B.

Further reading:

AWS Machine Learning documentation: https://aws.amazon.com/machine-learning/

Classification Accuracy: https://developers.google.com/machine-learning/crash-
course/classification/accuracy

Question: 5
A company is using a pre-trained large language model (LLM) to build a chatbot for product recommendations. The
company needs the LLM outputs to be short and written in a specific language.
Which solution will align the LLM response quality with the company's expectations?

A.Adjust the prompt.

B.Choose an LLM of a different size.
C.Increase the temperature.
D.Increase the Top K value.

Answer: A

Explanation:

The correct answer is A: Adjust the prompt. Here's a detailed justification:

The primary goal is to influence the LLM's output to be shorter and in a specific language. Prompt engineering
is the most direct and efficient way to achieve this. A well-crafted prompt acts as a precise instruction set,
guiding the LLM to generate the desired response. By including specific instructions within the prompt (e.g.,
"Answer in less than 20 words, in Spanish"), the company can directly shape the output to meet its
requirements.

Option B, choosing an LLM of a different size, is a less targeted approach. While different LLMs might have
varying tendencies, switching models introduces significant complexity and potential disruption without
directly addressing the specific needs of output length and language.

Option C, increasing the temperature, introduces more randomness and creativity into the output. This would
likely make the output less predictable and harder to control in terms of length and language adherence,
which is the opposite of the desired effect. Temperature controls the randomness of the token selection; a
higher temperature leads to more diverse and potentially less coherent responses.

Option D, increasing the Top K value, expands the pool of potential tokens the LLM considers before making a
prediction. While it can lead to more diverse outputs, it doesn't directly enforce length constraints or
language specifications. Like temperature, increasing Top K primarily affects diversity, not specific style or
formatting.

Therefore, prompt engineering offers the most granular and effective control over LLM output. Specific
prompts can be designed to directly address the needs of producing short, language-specific responses,
 making it the best solution in this scenario. Techniques such as few-shot prompting (providing examples of
desired output) can further refine the LLM's responses.

For further research:

Prompt Engineering Guide: https://www.promptingguide.ai/

Google AI - Prompt Design: https://developers.google.com/machine-learning/prompt-design (This link might
be outdated, but searches on "Google AI Prompt Design" will show the current similar pages.)

These resources will provide comprehensive information on crafting effective prompts for LLMs.

Question: 6
A company uses Amazon SageMaker for its ML pipeline in a production environment. The company has large input
data sizes up to 1 GB and processing times up to 1 hour. The company needs near real-time latency.
Which SageMaker inference option meets these requirements?

A.Real-time inference
B.Serverless inference
C.Asynchronous inference
D.Batch transform

Answer: C

Explanation:

Here's a detailed justification for why Asynchronous Inference (Option C) is the most suitable SageMaker
inference option for the company's requirements:

The company deals with input data sizes up to 1 GB and processing times extending to 1 hour while demanding
near real-time latency. This combination presents a unique challenge.

Real-time Inference (Option A): While designed for low latency, real-time inference typically handles smaller
payloads and shorter processing times. One-hour processing times are highly unusual and often impractical
for real-time endpoints. This could lead to timeouts and degraded performance as the endpoint needs to stay
active for a prolonged duration.

Serverless Inference (Option B): Serverless Inference automatically scales resources based on incoming
request traffic. While cost-effective for sporadic traffic, the 1 GB input size and 1-hour processing time could
potentially cause issues with Lambda function execution limits, if Serverless Inference relies on Lambda
under the hood. Furthermore, cold starts associated with serverless deployments may introduce
unacceptable latency for near real-time needs, especially for such lengthy inference times.

Asynchronous Inference (Option C): Asynchronous Inference is explicitly designed for handling large
payloads and long processing times without blocking the client. It decouples the request from the response,
allowing the client to submit a request and retrieve the results later. This makes it ideal for scenarios where
immediate responses are not crucial, but timely delivery of results is. The near real-time requirement can be
met if the processing time is generally within the acceptable range, and retrieval mechanism is efficient.
SageMaker manages queuing the requests, executing the model, and storing the output.

Batch Transform (Option D): Batch Transform is intended for offline inference on large datasets. It's not
suitable for near real-time latency requirements as the data is processed in batches, and there would be a
significant delay.

Therefore, Asynchronous Inference is the best fit because it supports large input sizes and long processing
 times, allowing the company to meet the near real-time latency requirement by decoupling the inference
request and enabling a timely output retrieval. The 1-hour processing time is a key factor favoring
asynchronous over real-time solutions.

Supporting Resources:

AWS SageMaker Asynchronous Inference: https://docs.aws.amazon.com/sagemaker/latest/dg/async-

inference.html
AWS SageMaker Inference Options: https://docs.aws.amazon.com/sagemaker/latest/dg/deploy-model.html

Question: 7
A company is using domain-specific models. The company wants to avoid creating new models from the beginning.
The company instead wants to adapt pre-trained models to create models for new, related tasks.
Which ML strategy meets these requirements?

A.Increase the number of epochs.

B.Use transfer learning.
C.Decrease the number of epochs.
D.Use unsupervised learning.

Answer: B

Explanation:

The correct answer is B. Use transfer learning.

Transfer learning is a machine learning technique where a model developed for a task is reused as the
starting point for a model on a second task. In this scenario, the company already possesses domain-specific
models. Transfer learning allows them to leverage the knowledge these pre-trained models have acquired,
rather than building entirely new models from scratch. By adapting these existing models to new, related
tasks, the company can significantly reduce training time, computational resources, and the amount of new
labeled data required. This is because the pre-trained model already understands relevant features and
patterns from its initial training. The pre-trained model's weights serve as a valuable initialization for the new
task, accelerating convergence and improving performance. Techniques like fine-tuning, feature extraction,
or a combination of both can be employed to adapt the pre-trained model.

Option A, increasing the number of epochs, only extends the training of the same model, not leveraging
knowledge from existing models. Option C, decreasing the number of epochs, would likely result in
underfitting. Option D, unsupervised learning, is generally used to discover patterns in unlabeled data, not to
adapt existing models for new tasks.

Further Research:

AWS Documentation on Transfer Learning: https://aws.amazon.com/machine-learning/transfer-learning/

(While not specific to AWS, it is a general ML concept AWS utilizes.)
Transfer Learning (Stanford CS231n): http://cs231n.github.io/transfer-learning/

Question: 8
A company is building a solution to generate images for protective eyewear. The solution must have high accuracy
and must minimize the risk of incorrect annotations.
Which solution will meet these requirements?
 A.Human-in-the-loop validation by using Amazon SageMaker Ground Truth Plus
B.Data augmentation by using an Amazon Bedrock knowledge base
C.Image recognition by using Amazon Rekognition
D.Data summarization by using Amazon QuickSight Q

Answer: A

Explanation:

Here's a detailed justification for why option A is the correct answer:

The core requirement is high accuracy and minimizing incorrect annotations for image generation related to
protective eyewear. This points to a need for meticulous review and correction of the generated images
and/or their training data.

Amazon SageMaker Ground Truth Plus (A): This service provides a managed workforce to perform data
labeling and validation. Its human-in-the-loop (HITL) capabilities directly address the accuracy requirement.
Trained workers can review generated images, correct annotations, and ensure the generated images align
with the desired outcome. This minimizes the risk of feeding flawed data into the image generation model,
consequently improving the overall quality and accuracy of the generated eyewear images. Ground Truth
Plus's managed workforce takes care of sourcing and managing annotators, and includes project
management tools to ensure high-quality labels are delivered.

Amazon Bedrock knowledge base (B): Data augmentation through a knowledge base in Bedrock primarily
focuses on expanding the dataset using existing knowledge. While it might improve the model's robustness, it
doesn't guarantee the correction of incorrect annotations in the first place. It's better for improving a model's
understanding of concepts than for ensuring initial annotation accuracy.

Amazon Rekognition (C): Rekognition is an image recognition service. While helpful for object detection after
image generation, it doesn't directly address the initial annotation or image generation quality control needed
to minimize errors during the process.

Amazon QuickSight (D): QuickSight is a data visualization and business intelligence tool. It provides insights
from data, but it doesn't contribute to image generation accuracy or annotation correction.

Therefore, the optimal choice is SageMaker Ground Truth Plus because it uniquely offers human validation to
minimize the risk of incorrect annotations during image generation and enhance accuracy.

Further Research:

Amazon SageMaker Ground Truth Plus: https://aws.amazon.com/sagemaker/groundtruth/

Amazon Bedrock: https://aws.amazon.com/bedrock/
Amazon Rekognition: https://aws.amazon.com/rekognition/
Amazon QuickSight: https://aws.amazon.com/quicksight/

Question: 9
A company wants to create a chatbot by using a foundation model (FM) on Amazon Bedrock. The FM needs to
access encrypted data that is stored in an Amazon S3 bucket. The data is encrypted with Amazon S3 managed
keys (SSE-S3).
The FM encounters a failure when attempting to access the S3 bucket data.
Which solution will meet these requirements?

A.Ensure that the role that Amazon Bedrock assumes has permission to decrypt data with the correct
encryption key.
B.Set the access permissions for the S3 buckets to allow public access to enable access over the internet.
 C.Use prompt engineering techniques to tell the model to look for information in Amazon S3.
D.Ensure that the S3 data does not contain sensitive information.

Answer: A

Explanation:

The correct answer is A. Here's a detailed justification:

Amazon Bedrock needs specific permissions to access and process data within an S3 bucket, especially when
that data is encrypted. Since the data is encrypted with SSE-S3 (Amazon S3 managed keys), Bedrock must
have the necessary authorization to decrypt the data before it can be used by the foundation model for
chatbot functionalities. This authorization is granted through an IAM role that Bedrock assumes when
interacting with other AWS services, including S3.

Option A directly addresses the root cause of the failure: lack of decryption permissions. By ensuring the IAM
role assigned to Bedrock has the s3:GetObject permission on the S3 bucket and the necessary permission to
decrypt the data using the SSE-S3 key, the foundation model can successfully access and utilize the
encrypted data. Specifically, the role needs kms:Decrypt permissions if using KMS managed keys, but SSE-S3
keys don't require explicit KMS permissions.

Option B is incorrect and a security risk. Granting public access to the S3 bucket exposes sensitive data to the
internet, violating the principle of least privilege and potentially leading to data breaches. It's a completely
unacceptable solution for handling encrypted data.

Option C is also incorrect. Prompt engineering is a technique used to refine the instructions given to a
foundation model to influence its output. It does not grant the model the necessary permissions to access and
decrypt data in S3. The model needs authorization to access the data before prompt engineering comes into
play.

Option D is related to data governance but does not solve the immediate problem of the foundation model
failing to access the encrypted data. Even if the data doesn't contain sensitive information, the model still
needs the required permissions to access and decrypt it. Furthermore, simply removing sensitive data doesn't
address the core requirement of using encrypted data for the chatbot. The data is encrypted for security, and
removing the sensitive information does not obviate the need for encryption.

In conclusion, the only way to fix the failure is to ensure Bedrock's IAM role has the appropriate s3:GetObject
permissions on the S3 bucket and access to decrypt the data using the key associated with SSE-S3. This
adheres to best practices for secure access to AWS resources.

References:

IAM Roles: https://docs.aws.amazon.com/IAM/latest/UserGuide/id_roles.html

Amazon S3 Encryption: https://docs.aws.amazon.com/AmazonS3/latest/userguide/serv-side-encryption.html
AWS Bedrock: https://aws.amazon.com/bedrock/

Question: 10
A company wants to use language models to create an application for inference on edge devices. The inference
must have the lowest latency possible.
Which solution will meet these requirements?

A.Deploy optimized small language models (SLMs) on edge devices.

B.Deploy optimized large language models (LLMs) on edge devices.
C.Incorporate a centralized small language model (SLM) API for asynchronous communication with edge
 devices.
D.Incorporate a centralized large language model (LLM) API for asynchronous communication with edge
devices.

Answer: A

Explanation:

The correct answer is A: Deploy optimized small language models (SLMs) on edge devices.

Here's why: The primary requirement is the lowest possible latency for inference. Latency refers to the delay
between a request and the response. Edge devices process data locally, reducing network hops and round-
trip times compared to centralized servers.

Options C and D involve centralized APIs, meaning data must travel to the server and back, adding significant
latency due to network communication. Asynchronous communication, while improving throughput in some
cases, does not directly address the need for lowest latency on a single inference request.

LLMs (Large Language Models), while powerful, are computationally intensive and require more resources to
run. Deploying them on edge devices would likely result in higher latency due to the processing constraints of
edge hardware.

SLMs (Small Language Models) are designed to be more efficient and can be optimized for edge deployment.
They require less computational power and memory, enabling faster inference on edge devices. By optimizing
the SLM, the company can further reduce latency. Deploying directly on the edge eliminates network latency,
fulfilling the requirement of the lowest possible latency.

Therefore, deploying optimized SLMs on edge devices is the most suitable approach for minimizing inference
latency in this scenario. This approach aligns with the principles of edge computing, where data processing is
moved closer to the source to reduce latency and bandwidth consumption.

For further research on Edge Computing and model deployment:

AWS Documentation on Edge Computing: https://aws.amazon.com/edge/

AWS SageMaker Edge Manager: https://aws.amazon.com/sagemaker/edge-manager/
Paperspace Blog on SLMs vs. LLMs: (search on Google for "paperspace small language model vs large
language model" - there isn't a definitive direct AWS link for this concept but many vendor sites will have
helpful comparisons)

Question: 11
A company wants to build an ML model by using Amazon SageMaker. The company needs to share and manage
variables for model development across multiple teams.
Which SageMaker feature meets these requirements?

A.Amazon SageMaker Feature Store

B.Amazon SageMaker Data Wrangler
C.Amazon SageMaker Clarify
D.Amazon SageMaker Model Cards

Answer: A

Explanation:

The correct answer is A. Amazon SageMaker Feature Store. Here's why:

 SageMaker Feature Store is a fully managed, centralized repository for storing, managing, and sharing ML
features. It enables organizations to define, store, and retrieve features in a consistent and scalable manner,
making them readily available for model training, inference, and feature exploration. For a company looking to
share and manage variables (features) across multiple teams for model development in SageMaker, Feature
Store provides the ideal solution.

Centralized Repository: Feature Store provides a central place to define and store features, preventing
feature duplication and ensuring consistency across teams.
Collaboration: Multiple teams can access and use the same features, fostering collaboration and reducing
redundant feature engineering efforts.
Versioning: Feature Store often supports versioning, allowing teams to track changes to features and ensure
reproducibility of models.
Discoverability: Feature Store provides a mechanism to discover and understand available features, making it
easier for teams to find the right features for their models.
Scalability: Feature Store is designed to handle large volumes of feature data and can scale to meet the
needs of enterprise-level ML projects.

Alternatives are not suited:

B. Amazon SageMaker Data Wrangler: primarily focuses on data preparation and feature engineering; it's not
designed for sharing and managing features across teams in the same way as Feature Store.
C. Amazon SageMaker Clarify: is used for bias detection and explainability of ML models, not for feature
sharing.
D. Amazon SageMaker Model Cards: is used to document and track information about ML models, but it
doesn't manage or share features.

In summary, SageMaker Feature Store directly addresses the requirement of sharing and managing variables
(features) for model development across multiple teams within the SageMaker environment. It facilitates
collaboration, ensures consistency, and simplifies feature management, making it the most suitable choice.

Relevant links:

Amazon SageMaker Feature Store: https://aws.amazon.com/sagemaker/feature-store/

SageMaker Feature Store documentation: https://docs.aws.amazon.com/sagemaker/latest/dg/feature-
store.html

Question: 12
A company wants to use generative AI to increase developer productivity and software development. The company
wants to use Amazon Q Developer.
What can Amazon Q Developer do to help the company meet these requirements?

A.Create software snippets, reference tracking, and open source license tracking.
B.Run an application without provisioning or managing servers.
C.Enable voice commands for coding and providing natural language search.
D.Convert audio files to text documents by using ML models.

Answer: A

Explanation:

The most suitable answer is A because Amazon Q Developer is specifically designed to boost developer
productivity by directly assisting with coding tasks. Option A accurately reflects this by listing key
functionalities like creating software snippets, which accelerates code development by providing ready-to-
 use code blocks. Reference tracking within Amazon Q helps developers understand the origin and context of
code elements, improving maintainability and reducing errors. Open-source license tracking is crucial for
compliance and avoiding legal issues when using open-source components. These features directly address
the company's goal of enhancing developer productivity and software development.

Option B describes serverless computing, which is a characteristic of services like AWS Lambda, and while
beneficial for certain applications, it doesn't directly contribute to enhanced coding assistance provided by
Amazon Q Developer.

Option C, enabling voice commands for coding, is an interesting capability but is not the core focus or primary
function advertised for Amazon Q Developer. While natural language search might be incorporated, the main
thrust is on code generation and understanding.

Option D pertains to audio transcription, a function handled by services like Amazon Transcribe, and is
unrelated to the core objective of increasing developer productivity through code assistance.

Therefore, given the company's need to increase developer productivity and software development using
generative AI, the most relevant answer is A.For more information, research Amazon Q Developer on the AWS
website: https://aws.amazon.com/q/developer/

Question: 13
A financial institution is using Amazon Bedrock to develop an AI application. The application is hosted in a VPC. To
meet regulatory compliance standards, the VPC is not allowed access to any internet traffic.
Which AWS service or feature will meet these requirements?

A.AWS PrivateLink
B.Amazon Macie
C.Amazon CloudFront
D.Internet gateway

Answer: A

Explanation:

The correct answer is A. AWS PrivateLink. Here's a detailed justification:

AWS PrivateLink enables you to access AWS services and services hosted by other AWS accounts (referred
to as endpoint services) in a private and secure manner, without exposing your traffic to the public internet.
This is achieved by establishing private connectivity between your VPC and the service using Elastic Network
Interfaces (ENIs) within your VPC.

In the scenario described, the financial institution requires a secure connection to Amazon Bedrock within a
VPC that doesn't allow internet access due to regulatory compliance. AWS PrivateLink directly addresses this
requirement. By creating a VPC endpoint for Amazon Bedrock powered by PrivateLink, the application within
the VPC can privately access Bedrock's API without traversing the internet. This connection is isolated within
the AWS network.

Let's examine why the other options are incorrect:

B. Amazon Macie: Macie is a data security and data privacy service that uses machine learning and pattern
matching to discover and protect sensitive data in AWS. It doesn't establish private connectivity.

C. Amazon CloudFront: CloudFront is a content delivery network (CDN) used to distribute content with low
latency and high transfer speeds. It typically involves internet access and isn't suitable for completely
 isolating traffic within a VPC.

D. Internet Gateway: An Internet Gateway allows resources within a VPC to access the internet. This directly
contradicts the requirement of no internet access.

In summary, AWS PrivateLink provides the necessary private connectivity to Amazon Bedrock from within the
restricted VPC, fulfilling the regulatory compliance requirements of the financial institution.

Authoritative Links:

AWS PrivateLink Documentation

Amazon Bedrock Documentation

Question: 14
A company wants to develop an educational game where users answer questions such as the following: "A jar
contains six red, four green, and three yellow marbles. What is the probability of choosing a green marble from the
jar?"
Which solution meets these requirements with the LEAST operational overhead?

A.Use supervised learning to create a regression model that will predict probability.
B.Use reinforcement learning to train a model to return the probability.
C.Use code that will calculate probability by using simple rules and computations.
D.Use unsupervised learning to create a model that will estimate probability density.

Answer: C

Explanation:

The correct answer is C. Use code that will calculate probability by using simple rules and computations.

Justification:

The problem describes a scenario where probability calculation is straightforward and based on well-defined
mathematical formulas. The question explicitly involves counting and applying basic probability rules (number
of favorable outcomes divided by total number of outcomes). Implementing this calculation through code (e.g.,
Python) directly addresses the problem efficiently and accurately.

Operational Overhead: Using code for direct calculation introduces the least operational overhead. It requires
minimal infrastructure, no model training or deployment, and is computationally inexpensive.
Supervised Learning (Option A): Supervised learning would require a large dataset of questions and correct
probabilities to train a regression model. This introduces significant overhead for data collection, labeling,
model training, and deployment. It is an over-engineered solution for a simple problem.
Reinforcement Learning (Option B): Reinforcement learning would involve training an agent to answer
probability questions. This is highly inappropriate and inefficient for a problem with a deterministic solution.
It's complex to implement and requires extensive training and reward engineering.
Unsupervised Learning (Option D): Unsupervised learning techniques like probability density estimation are
not relevant here. The task is to compute a specific probability, not to understand the underlying distribution
of probability values.

Since there is already a formula that calculates probability, utilizing it would lead to the least overhead. Using
machine learning would increase complexity while providing no extra benefit.

In cloud computing, minimizing operational overhead is a key design principle. By using a simple code-based
solution, the company reduces infrastructure costs, maintenance efforts, and overall complexity while still
meeting the requirements of the educational game.
 Here are some authoritative links for further research:

Basic Probability: https://www.mathsisfun.com/data/probability.html

Calculating Probability: https://www.khanacademy.org/math/statistics-probability/counting-permutations-
and-combinations

Question: 15
Which metric measures the runtime efficiency of operating AI models?

A.Customer satisfaction score (CSAT)

B.Training time for each epoch
C.Average response time
D.Number of training instances

Answer: C

Explanation:

Here's a detailed justification for why average response time is the best metric for measuring the runtime
efficiency of operating AI models:

Average response time directly reflects how quickly an AI model provides a prediction or output in a real-
world application. It's a crucial indicator of the model's operational performance from a user's perspective. A
shorter average response time implies that the model is processing requests efficiently, leading to a better
user experience. This is especially vital in latency-sensitive applications like real-time recommendations,
fraud detection, or conversational AI, where delays can negatively impact usability and effectiveness.

While training time (option B) is important during model development, it doesn't directly measure runtime
efficiency. The training phase focuses on learning patterns from data, while runtime refers to how quickly the
model provides predictions after it has been deployed. Customer satisfaction (CSAT), option A, is a broad
measure of user experience but can be affected by numerous factors beyond the model's runtime
performance. Finally, option D, "Number of training instances," pertains to the dataset used for training and is
not a metric for runtime efficiency.

The principle of minimizing latency is fundamental in cloud computing and AI deployment. Efficient models
contribute to lower operational costs by consuming fewer resources and minimizing the need for scaling
infrastructure. Average response time allows DevOps and MLOps engineers to monitor model performance,
identify bottlenecks, and optimize resource allocation. Slow response times could indicate issues such as
inefficient model code, insufficient compute resources, or network latency problems.

In cloud environments like AWS, services such as Amazon CloudWatch can be used to monitor and alarm on
average response time metrics for deployed AI models. Techniques like model optimization, caching, and the
selection of appropriate instance types can be employed to improve response times.

Further research on model deployment and performance monitoring on AWS can be found at the following
links:

AWS Documentation on Monitoring Machine Learning Models:

https://docs.aws.amazon.com/sagemaker/latest/dg/model-monitor.html
AWS Documentation on CloudWatch: https://docs.aws.amazon.com/cloudwatch/

Question: 16
A company is building a contact center application and wants to gain insights from customer conversations. The
company wants to analyze and extract key information from the audio of the customer calls.
Which solution meets these requirements?

A.Build a conversational chatbot by using Amazon Lex.

B.Transcribe call recordings by using Amazon Transcribe.
C.Extract information from call recordings by using Amazon SageMaker Model Monitor.
D.Create classification labels by using Amazon Comprehend.

Answer: B

Explanation:

The correct answer is B. Transcribe call recordings by using Amazon Transcribe.

Here's a detailed justification:

The core requirement is to gain insights and extract key information from customer conversation audio.
Amazon Transcribe directly addresses this by converting audio into text. This transcription process allows for
subsequent analysis to identify key phrases, sentiment, topics discussed, and other relevant data points.

Option A (Amazon Lex) is more suitable for building conversational interfaces, not directly analyzing existing
audio recordings. While Lex could be integrated later to process the transcribed text, it's not the initial step to
extract information from the audio itself.

Option C (Amazon SageMaker Model Monitor) is for monitoring the performance of machine learning models,
not for directly transcribing or analyzing audio. It's a downstream tool that might be relevant after analysis,
but not for the primary task of extracting information from the calls.

Option D (Amazon Comprehend) performs natural language processing (NLP) tasks like sentiment analysis
and entity recognition. Comprehend is useful, but it requires text as input. It cannot directly process audio; it
needs transcribed text from a service like Transcribe.

Therefore, Amazon Transcribe is the most logical first step as it bridges the gap between audio data and text-
based analysis tools like Comprehend. By transcribing the call recordings, the company will then be able to
leverage NLP tools or even manual analysis to glean the desired insights from the conversations.

Essentially, Transcribe provides the textual data necessary for subsequent analysis to uncover valuable
information from the calls.

Authoritative Links:

Amazon Transcribe: https://aws.amazon.com/transcribe/

Amazon Comprehend: https://aws.amazon.com/comprehend/

Question: 17
A company has petabytes of unlabeled customer data to use for an advertisement campaign. The company wants
to classify its customers into tiers to advertise and promote the company's products.
Which methodology should the company use to meet these requirements?

A.Supervised learning
B.Unsupervised learning
C.Reinforcement learning
D.Reinforcement learning from human feedback (RLHF)
 Answer: B

Explanation:

The company needs to classify its customers into tiers using petabytes of unlabeled data. This is a classic
clustering problem, where the goal is to group similar data points together without any prior knowledge of the
correct labels.

Supervised learning (Option A) requires labeled data to train a model. Since the data is unlabeled, supervised
learning is not applicable. Supervised learning algorithms learn a mapping function from input features to
output labels using labeled training data.

Unsupervised learning (Option B) is ideal for this scenario. Unsupervised learning algorithms discover patterns
and structure in unlabeled data. Clustering algorithms, a type of unsupervised learning, can automatically
group customers into tiers based on their inherent similarities, without requiring pre-defined labels. Common
unsupervised learning methods include k-means clustering, hierarchical clustering, and DBSCAN. The
company can use algorithms like k-means to automatically segment the customers based on their
characteristics derived from the data.

Reinforcement learning (Option C) involves training an agent to make decisions in an environment to maximize
a reward. It's not relevant for classifying data into tiers. Reinforcement learning is typically applied when an
agent needs to learn optimal actions through trial and error in an environment.

Reinforcement learning from human feedback (RLHF) (Option D) is an advanced technique within
reinforcement learning where human feedback is used to guide the agent's learning process. It's irrelevant
because the task does not involve an agent and rewards.

Therefore, unsupervised learning (Option B) is the correct methodology because it allows the company to
discover inherent groupings within their unlabeled customer data and classify them into tiers, thereby
meeting the requirements for their advertisement campaign.

Further Reading:

AWS AI Services: https://aws.amazon.com/ai/

Unsupervised Learning on AWS: https://aws.amazon.com/blogs/machine-learning/performing-a-k-means-
clustering-analysis-with-amazon-athena/

Question: 18
An AI practitioner wants to use a foundation model (FM) to design a search application. The search application
must handle queries that have text and images.
Which type of FM should the AI practitioner use to power the search application?

A.Multi-modal embedding model

B.Text embedding model
C.Multi-modal generation model
D.Image generation model

Answer: A

Explanation:

The AI practitioner needs a model that can understand and compare both text and images to power the search
application. A multi-modal embedding model is the most appropriate choice.

Here's why:
 Multi-modal: The application requires handling two different modalities: text and images. A multi-modal
model is designed to process and relate information from different modalities.
Embedding: Embedding models create vector representations of the input data (text and images in this case).
These vector representations capture the semantic meaning of the data.
Search Application: The core of a search application is comparing queries to the indexed content. Embedding
models allow for efficient similarity search in a vector space. The text and image queries can be converted
into embeddings and compared against embeddings of the indexed data.
Comparison: By embedding both text and images into a common vector space, the model allows for the
comparison of text queries to images and vice versa. The system can then retrieve the most relevant content
based on the similarity of their embeddings.
Alternatives are unsuitable: Text embedding models only handle text, and image generation models create
new images rather than find relevant ones. Multi-modal generation models create new content from different
inputs, whereas, the search application requires comparing existing content.

In contrast, a text embedding model would only handle text queries, an image generation model would create
new images rather than find relevant ones, and a multi-modal generation model would create new text or
images based on the inputs rather than identify similarity and relevance.

Further Reading:

Amazon Bedrock Multi-Modal Embeddings: https://aws.amazon.com/bedrock/multi-modal-embeddings/

(This link talks about multi-modal capabilities in Amazon Bedrock, illustrating the concept).
Multi-Modal Learning: https://www.cs.cmu.edu/~mmv/ (Carnegie Mellon University's research on multi-modal
learning.)

Question: 19
A company uses a foundation model (FM) from Amazon Bedrock for an AI search tool. The company wants to fine-
tune the model to be more accurate by using the company's data.
Which strategy will successfully fine-tune the model?

A.Provide labeled data with the prompt field and the completion field.
B.Prepare the training dataset by creating a .txt file that contains multiple lines in .csv format.
C.Purchase Provisioned Throughput for Amazon Bedrock.
D.Train the model on journals and textbooks.

Answer: A

Explanation:

The correct answer is A. Provide labeled data with the prompt field and the completion field.

Fine-tuning a foundation model (FM) in Amazon Bedrock involves adapting the pre-trained model to perform
better on a specific task using your own data. This is achieved by providing the model with examples of inputs
(prompts) and the desired outputs (completions). This teaches the model to generate more accurate and
relevant responses for your use case.

Option A directly addresses this fine-tuning process. By providing labeled data in a format of prompt-
completion pairs, you are guiding the model to learn the desired relationships and improve its accuracy on
your specific data. The prompt acts as the input to the model, and the completion represents the ideal output
for that prompt. This is a common method of fine-tuning large language models.

Option B is incorrect. While a .txt file with .csv format might be used for data storage, it doesn't inherently
define the prompt-completion structure required for fine-tuning. The data must be specifically formatted to
indicate which part is the input and which is the desired output.
 Option C is irrelevant. Provisioned Throughput in Amazon Bedrock relates to ensuring dedicated capacity and
performance for inference, not to the fine-tuning process itself. It addresses the speed and availability of
predictions from a model but does not contribute to its accuracy.

Option D is too generic. While training on journals and textbooks could improve a model's general knowledge,
it doesn't guarantee improved performance on the company's specific AI search task. Fine-tuning with
relevant, labeled data from the company is far more effective for achieving the desired accuracy.

In summary, the prompt-completion format enables the model to learn the desired relationships within your
data, resulting in a fine-tuned model that is more accurate for the AI search tool.

Relevant links for further research:

Amazon Bedrock documentation

Fine-tuning Large Language Models

Question: 20
A company wants to use AI to protect its application from threats. The AI solution needs to check if an IP address is
from a suspicious source.
Which solution meets these requirements?

A.Build a speech recognition system.

B.Create a natural language processing (NLP) named entity recognition system.
C.Develop an anomaly detection system.
D.Create a fraud forecasting system.

Answer: C

Explanation:

The correct answer is C, developing an anomaly detection system. Here's why:

The problem describes a scenario requiring the identification of suspicious IP addresses accessing an
application. Anomaly detection, as a branch of AI, is specifically designed to identify deviations from normal
behavior patterns. In this context, "normal" would represent typical IP address access patterns, and
"anomalous" would be unusual IPs indicating potential threats. AWS offers services like Amazon GuardDuty
and Amazon CloudWatch Anomaly Detection that can be leveraged for this purpose. GuardDuty, for example,
analyzes VPC Flow Logs, DNS logs, and CloudTrail logs to identify malicious or unauthorized behavior.

Options A, B, and D are incorrect because they address different AI problems. Speech recognition (A) focuses
on converting audio into text, irrelevant to network security. NLP named entity recognition (B) identifies and
categorizes entities (e.g., people, organizations) in text, also unrelated to IP address analysis. Fraud
forecasting (D) predicts future fraudulent activities, whereas the requirement is to detect current suspicious
activity, making anomaly detection the more appropriate solution.

Anomaly detection systems can be trained on historical network traffic data to establish a baseline of normal
activity. When new IP addresses access the application, the system compares their behavior against this
baseline. If an IP address exhibits unusual access patterns (e.g., accessing resources at unusual times,
generating an abnormally high number of requests), the system flags it as suspicious. The detected anomalies
can trigger alerts or automated responses, such as blocking the IP address.

Supporting Links:

Amazon GuardDuty: https://aws.amazon.com/guardduty/

Amazon CloudWatch Anomaly Detection: https://aws.amazon.com/cloudwatch/features/anomaly-detection/
 Anomaly detection: https://en.wikipedia.org/wiki/Anomaly_detection

Question: 21
Which feature of Amazon OpenSearch Service gives companies the ability to build vector database applications?

A.Integration with Amazon S3 for object storage

B.Support for geospatial indexing and queries
C.Scalable index management and nearest neighbor search capability
D.Ability to perform real-time analysis on streaming data

Answer: C

Explanation:

The correct answer is C: Scalable index management and nearest neighbor search capability. This is because
vector databases excel at storing and querying high-dimensional vector embeddings, which represent data
points in a semantic space. Amazon OpenSearch Service's ability to manage indexes at scale is crucial for
handling the large datasets typically associated with vector embeddings. More importantly, the "nearest
neighbor search capability" allows for efficient similarity searches, finding vectors closest to a query vector,
which is the fundamental operation in vector database applications like semantic search, recommendation
systems, and image recognition. Options A, B, and D, while valid features of OpenSearch Service, are not
specific to the needs of vector database applications. A relates to general object storage, B to location-based
data, and D to real-time data analysis. While real-time analysis can leverage vector embeddings, the core
function is provided by the nearest neighbor search and scalable index management. Feature stores leverage
vector embeddings; therefore, the ability to efficiently search and scale is critical. A vector database built
upon OpenSearch needs those capabilities directly.

For further research, explore the following resources:

Amazon OpenSearch Service Documentation: https://aws.amazon.com/opensearch-service/

Amazon OpenSearch Service k-NN: https://opensearch.org/docs/latest/search-plugins/knn/index/
AWS AI and Machine Learning: https://aws.amazon.com/ai/

Question: 22
Which option is a use case for generative AI models?

A.Improving network security by using intrusion detection systems

B.Creating photorealistic images from text descriptions for digital marketing
C.Enhancing database performance by using optimized indexing
D.Analyzing financial data to forecast stock market trends

Answer: B

Explanation:

The correct answer is B, creating photorealistic images from text descriptions for digital marketing.
Generative AI models are designed to generate new, original content. Option B directly aligns with this core
function. Text-to-image models, a specific type of generative AI, excel at producing visual content based on
textual prompts. This makes them highly valuable for digital marketing, where visually appealing and unique
images can enhance campaigns and attract customers.
 Option A, improving network security with intrusion detection systems, relates to AI in cybersecurity,
particularly anomaly detection. While AI is used, this doesn't necessarily involve generating content. Option C,
enhancing database performance with optimized indexing, employs AI for optimization, a different application
area. Option D, analyzing financial data for stock market trends, falls under predictive analytics, again, an area
separate from generative AI's content creation focus.

Generative AI models like DALL-E 2, Stable Diffusion, and Midjourney showcase the ability to produce
photorealistic images, artwork, and variations based on text prompts. These models are transforming digital
marketing by allowing users to create customized visuals without needing traditional photography or design
skills.

For more information on Generative AI and its applications:

AWS Documentation on Generative AI: (Search AWS Documentation for "Generative AI" or "Amazon
Bedrock" for their offerings)
Google AI Blog on Generative Models: https://ai.googleblog.com/ (Search for "Generative Models")
OpenAI's DALL-E 2: https://openai.com/dall-e-2/

Question: 23
A company wants to build a generative AI application by using Amazon Bedrock and needs to choose a foundation
model (FM). The company wants to know how much information can fit into one prompt.
Which consideration will inform the company's decision?

A.Temperature
B.Context window
C.Batch size
D.Model size

Answer: B

Explanation:

The correct answer is B. Context window.

The context window of a foundation model (FM) in Amazon Bedrock dictates the maximum amount of text that
can be included within a single prompt and its associated response. This is a critical consideration for a
generative AI application because it directly impacts the amount of information the model can effectively
process and utilize to generate relevant and coherent outputs.

A larger context window allows the model to consider more data, enabling it to handle longer documents,
maintain better context in conversations, and generate more detailed and nuanced responses. Conversely, a
smaller context window limits the model's ability to leverage extensive information, potentially leading to less
accurate or complete results.

Why the other options are incorrect:

A. Temperature: Temperature controls the randomness of the model's output. While important for controlling
creativity and predictability, it doesn't affect how much data can be fed into the model.
C. Batch size: Batch size refers to the number of prompts processed simultaneously. This is relevant for
throughput and efficiency, but not the size of a single prompt.
D. Model size: Model size, referring to the number of parameters in the FM, indicates the model's complexity
and potential performance. However, it does not directly define the amount of text the model can process
within a prompt.
 Therefore, when selecting an FM for a generative AI application in Amazon Bedrock, the company must
consider the context window of each model to ensure it can accommodate the expected input size for their
use case.

Supporting Links:

Amazon Bedrock Documentation: provides comprehensive details regarding the models and their capabilities.
Understanding LLM Context Windows: External resources for understanding context windows.

Question: 24
A company wants to make a chatbot to help customers. The chatbot will help solve technical problems without
human intervention.
The company chose a foundation model (FM) for the chatbot. The chatbot needs to produce responses that adhere
to company tone.
Which solution meets these requirements?

A.Set a low limit on the number of tokens the FM can produce.

B.Use batch inferencing to process detailed responses.
C.Experiment and refine the prompt until the FM produces the desired responses.
D.Define a higher number for the temperature parameter.

Answer: C

Explanation:

The correct answer is C: Experiment and refine the prompt until the FM produces the desired responses. This
is because prompt engineering is a key technique for tailoring the output of a foundation model to specific
requirements and stylistic guidelines.

Here's a detailed justification:

Foundation models are pre-trained on vast datasets and, while powerful, may not inherently understand or
adhere to a specific company tone or desired response style. Fine-tuning a pre-trained FM can be a heavy lift
in time and resources. Therefore, prompt engineering becomes a crucial, more cost-effective method for
influencing the model's output.

By carefully crafting prompts, which are the instructions given to the model, the company can guide the
chatbot to generate responses that align with their brand voice and target audience. Experimenting with
different prompt structures, keywords, and examples helps to discover the prompts that consistently produce
the desired responses. This iterative process of experimentation and refinement allows the company to fine-
tune the chatbot's behavior without directly retraining the FM.

Option A, setting a low token limit, primarily controls the length of the response, not the tone or style. While a
shorter response might indirectly influence the tone, it doesn't guarantee adherence to company guidelines.
Option B, using batch inferencing, is more related to processing large volumes of requests offline rather than
influencing the response style. Option D, increasing the temperature parameter, introduces more randomness
and creativity into the responses. While this might sometimes be desirable, it's generally not suitable for a
chatbot that needs to adhere to a specific, controlled tone and provide reliable information. A high
temperature can lead to unpredictable and inconsistent outputs, potentially deviating from the desired
company tone.

In conclusion, prompt engineering provides a direct and effective mechanism to guide the FM towards
producing responses that meet the specific requirements of the company's chatbot, making it the optimal
solution in this scenario.
 Relevant links:

Prompt Engineering Guide: https://www.promptingguide.ai/

Amazon Bedrock Documentation on Prompt Engineering: (Check AWS Documentation directly as links
expire)

Question: 25
A company wants to use a large language model (LLM) on Amazon Bedrock for sentiment analysis. The company
wants to classify the sentiment of text passages as positive or negative.

Which prompt engineering strategy meets these requirements?

A.Provide examples of text passages with corresponding positive or negative labels in the prompt followed by
the new text passage to be classified.
B.Provide a detailed explanation of sentiment analysis and how LLMs work in the prompt.
C.Provide the new text passage to be classified without any additional context or examples.
D.Provide the new text passage with a few examples of unrelated tasks, such as text summarization or question
answering.

Answer: A

Explanation:

The correct answer is A, providing examples of text passages with corresponding sentiment labels in the
prompt. This approach leverages a prompt engineering strategy known as "few-shot learning." Large
Language Models (LLMs) like those available through Amazon Bedrock excel when given examples of the
desired task. By including examples of text and their corresponding sentiment (positive or negative), you're
effectively demonstrating to the LLM how you want it to classify the new, unlabeled text.

This method guides the LLM to understand the nuances of positive and negative sentiment within the specific
context of the company's data. It helps the LLM generalize from these examples to the new input, improving
the accuracy and relevance of the sentiment classification. Options B, C, and D are less effective because they
don't provide the LLM with the specific examples needed for the targeted sentiment analysis task.

Providing a detailed explanation of sentiment analysis (Option B) is unnecessary, as LLMs are pre-trained on
vast amounts of text data and already possess a general understanding of the concept. Simply providing the
text passage without context (Option C) relies entirely on the LLM's inherent knowledge and might not be
sufficient for accurate sentiment classification in a specific domain or style. Presenting unrelated tasks
(Option D) would confuse the LLM and hinder its ability to focus on the sentiment analysis goal. Few-shot
learning gives the model a tangible and tailored guide to perform the specific task, leading to better results.
Few-shot learning optimizes accuracy by presenting the model with relevant examples, which enables the
model to adapt its responses to match the user's specific
needs.https://aws.amazon.com/bedrock/https://towardsdatascience.com/prompt-engineering-guide-for-
developers-a49d27609e88

Question: 26
A security company is using Amazon Bedrock to run foundation models (FMs). The company wants to ensure that
only authorized users invoke the models. The company needs to identify any unauthorized access attempts to set
appropriate AWS Identity and Access Management (IAM) policies and roles for future iterations of the FMs.
Which AWS service should the company use to identify unauthorized users that are trying to access Amazon
Bedrock?
 A.AWS Audit Manager
B.AWS CloudTrail
C.Amazon Fraud Detector
D.AWS Trusted Advisor

Answer: B

Explanation:

The correct answer is B. AWS CloudTrail.

AWS CloudTrail is a service that enables governance, compliance, operational auditing, and risk auditing of
your AWS account. CloudTrail logs API calls made to AWS services, including Amazon Bedrock. This means
that every attempt to invoke a Bedrock model, whether successful or unsuccessful, is recorded in CloudTrail
logs. These logs capture information about the identity of the caller (the IAM user or role used to make the
request), the time of the request, the source IP address, the specific API call made (e.g., InvokeModel), and
whether the request was authorized.

By analyzing these CloudTrail logs, the security company can identify unauthorized access attempts. They
can filter the logs for events where the IAM user or role attempting to access Amazon Bedrock does not have
the necessary permissions. This information is crucial for identifying users who are trying to access the
models without proper authorization. They can then refine their IAM policies and roles to prevent future
unauthorized access, ensuring only authorized users can invoke the models.

AWS Audit Manager (A) helps you continuously audit your AWS usage to simplify how you assess risk and
compliance with regulations and industry standards. While useful for compliance, it doesn't directly identify
specific unauthorized access attempts like CloudTrail. Amazon Fraud Detector (C) is used to detect fraudulent
activities and is not designed for tracking API access or unauthorized access attempts to AWS services like
Bedrock. AWS Trusted Advisor (D) provides recommendations for cost optimization, performance, security,
fault tolerance, and service limits. It does not track API calls or provide detailed information on unauthorized
access attempts.

In summary, CloudTrail is the appropriate service because it specifically logs API calls, including unauthorized
attempts, allowing the company to identify the users and the actions they were trying to perform, directly
addressing the need to identify unauthorized access attempts.

Here are some authoritative links for further research:

AWS CloudTrail: https://aws.amazon.com/cloudtrail/

Logging Amazon Bedrock API calls with AWS CloudTrail:
https://docs.aws.amazon.com/bedrock/latest/userguide/security-logging-using-cloudtrail.html

Question: 27
A company has developed an ML model for image classification. The company wants to deploy the model to
production so that a web application can use the model.
The company needs to implement a solution to host the model and serve predictions without managing any of the
underlying infrastructure.
Which solution will meet these requirements?

A.Use Amazon SageMaker Serverless Inference to deploy the model.

B.Use Amazon CloudFront to deploy the model.
C.Use Amazon API Gateway to host the model and serve predictions.
D.Use AWS Batch to host the model and serve predictions.
 Answer: A

Explanation:

The correct answer is A. Use Amazon SageMaker Serverless Inference to deploy the model.

Here's why:

Amazon SageMaker Serverless Inference is specifically designed to deploy machine learning models for
inference without managing underlying infrastructure. It automatically provisions and scales compute
resources based on the request volume, eliminating the need for manual capacity planning and management.
This aligns perfectly with the company's requirement to host the model and serve predictions without
infrastructure overhead. The model is invoked through an endpoint, making it suitable for integration with a
web application.

Option B, Amazon CloudFront, is a content delivery network (CDN) used for caching and distributing static and
dynamic web content. It's not designed for hosting and serving ML model predictions.

Option C, Amazon API Gateway, is used to create, publish, maintain, monitor, and secure APIs. While it can be
used as a front-end for accessing an ML model, it doesn't host or manage the model itself. You would still
need a compute resource behind API Gateway to serve the predictions.

Option D, AWS Batch, is a batch computing service that allows you to run batch computing workloads at any
scale. It's designed for running discrete jobs and isn't suitable for real-time, low-latency inference required by
a web application.

SageMaker Serverless Inference is the best option because it provides a fully managed environment for
hosting and serving ML models, automatically scaling to meet demand, and abstracting away the complexity
of infrastructure management.

Relevant links:

Amazon SageMaker Serverless Inference

Question: 28
An AI company periodically evaluates its systems and processes with the help of independent software vendors
(ISVs). The company needs to receive email message notifications when an ISV's compliance reports become
available.
Which AWS service can the company use to meet this requirement?

A.AWS Audit Manager

B.AWS Artifact
C.AWS Trusted Advisor
D.AWS Data Exchange

Answer: B

Explanation:

The correct answer is B (AWS Artifact). Here's why:

AWS Artifact is a service that provides on-demand access to AWS's compliance reports and agreements.
Critically, it also provides a mechanism for customers to download compliance reports from third-party
vendors who have chosen to share them through Artifact. AWS Artifact allows you to subscribe to
notifications when new reports are available. This directly addresses the requirement of receiving email
notifications when an ISV's compliance reports become available.
 AWS Audit Manager automates the process of auditing your AWS usage and provides evidence to support
audits. While important for compliance, it doesn't directly manage or distribute third-party compliance
reports, nor does it provide notifications about their availability.

AWS Trusted Advisor provides recommendations to optimize your AWS infrastructure for cost, performance,
security, and fault tolerance. It does not provide access to or notifications about third-party compliance
reports.

AWS Data Exchange is a service for finding, subscribing to, and using third-party data in the cloud. While it
deals with third-party data, it's focused on data sets for analysis and use, not compliance reports.

Therefore, AWS Artifact is the only service that specifically facilitates accessing and being notified about the
availability of third-party compliance reports, fulfilling the exam question's requirement.

For further information, refer to the official AWS Artifact documentation: https://aws.amazon.com/artifact/
and the AWS Audit Manager documentation https://aws.amazon.com/audit-manager/. Also refer to AWS
Trusted Advisor https://aws.amazon.com/premiumsupport/technology/trusted-advisor/ and AWS Data
Exchange documentation https://aws.amazon.com/data-exchange/.

Question: 29
A company wants to use a large language model (LLM) to develop a conversational agent. The company needs to
prevent the LLM from being manipulated with common prompt engineering techniques to perform undesirable
actions or expose sensitive information.
Which action will reduce these risks?

A.Create a prompt template that teaches the LLM to detect attack patterns.
B.Increase the temperature parameter on invocation requests to the LLM.
C.Avoid using LLMs that are not listed in Amazon SageMaker.
D.Decrease the number of input tokens on invocations of the LLM.

Answer: A

Explanation:

The correct answer is A: Create a prompt template that teaches the LLM to detect attack patterns.

Here's a detailed justification:

Large Language Models (LLMs) are susceptible to prompt injection attacks, where malicious users manipulate
the input to bypass intended security measures or elicit undesirable responses. A robust defense strategy
involves proactively training the LLM to recognize and neutralize these attack patterns. This is achieved by
crafting prompt templates that explicitly instruct the model on how to identify, categorize, and respond to
potential attacks. Such templates can include examples of common attack vectors like prompt leaking,
denial-of-service, and jailbreaking attempts.

By teaching the LLM to detect and respond to such patterns, you create a defensive layer that mitigates the
risk of the model being manipulated. This proactive approach is more effective than simply relying on input
sanitization or rate limiting. Prompt templates allow you to define boundaries for the LLM's behavior and
instruct it to reject or flag prompts that deviate from intended use cases.

Option B is incorrect because increasing the temperature parameter increases the randomness and creativity
of the LLM's responses, potentially making it more vulnerable to manipulation and generating unpredictable
or harmful outputs.
 Option C is incorrect because limiting the LLMs to only those on Amazon SageMaker does not inherently
protect from prompt injection. The security of the LLM ultimately depends on how it is configured and
secured, not solely on its availability within a particular platform.

Option D is incorrect because decreasing the number of input tokens may limit some complex attacks, but it
also restricts the LLM's ability to understand the user's intent and perform its intended function. This
approach is also insufficient to guard against all attack types and would negatively affect the usability of the
conversational agent.

Therefore, creating a prompt template that teaches the LLM to identify and neutralize attack patterns offers
the most effective means of mitigating the risks of prompt injection and ensuring that the conversational
agent operates within defined security boundaries.

Supporting Links:

Prompt Injection: https://owasp.org/www-project-top-ten-for-llm-applications/

Prompt Engineering: https://www.promptingguide.ai/

Question: 30
A company is using the Generative AI Security Scoping Matrix to assess security responsibilities for its solutions.
The company has identified four different solution scopes based on the matrix.
Which solution scope gives the company the MOST ownership of security responsibilities?

A.Using a third-party enterprise application that has embedded generative AI features.

B.Building an application by using an existing third-party generative AI foundation model (FM).
C.Refining an existing third-party generative AI foundation model (FM) by fine-tuning the model by using data
specific to the business.
D.Building and training a generative AI model from scratch by using specific data that a customer owns.

Answer: D

Explanation:

The answer is D because it represents the scenario where the company has the most control and therefore the
most responsibility for security. The Generative AI Security Scoping Matrix, which aims to clarify security
responsibilities across different engagement levels, would classify building and training a model from scratch
as requiring the most ownership.

Option A involves using a third-party application where the vendor largely manages the security of the
embedded AI features. Option B, building an application using an existing FM, shifts more security
responsibility to the company but still relies on the FM provider's security measures. Option C, fine-tuning an
FM, increases the company's security responsibility over the data used for fine-tuning, but the underlying
FM's security remains primarily the vendor's concern.

However, in Option D, the company controls every aspect of the model's lifecycle, including data ingestion,
model training, infrastructure security, and monitoring. This complete control directly translates to complete
accountability for ensuring the model's security and preventing vulnerabilities like data poisoning, model
evasion, or unintended bias. The entire security burden falls squarely on the company, as they're building the
system from the ground up. This includes not only the model's code and architecture, but also the security of
the data used for training and inference.

Therefore, the solution scope that necessitates the most ownership of security responsibilities is building and
training a generative AI model from scratch using specific data that the customer owns. This "from scratch"
approach means no reliance on external vendors for the core model's security, increasing the company's
 responsibilities significantly.

For further research, consider these resources:

OWASP (Open Web Application Security Project): Provides guidance on AI security risks and mitigations.
https://owasp.org/
NIST AI Risk Management Framework: Offers a framework for managing risks associated with AI systems.
https://www.nist.gov/itl/ai-risk-management-framework
ENISA (European Union Agency for Cybersecurity) - AI Cybersecurity: provides analysis of AI specific
cybersecurity threats and vulnerabilities https://www.enisa.europa.eu/topics/emerging-
technologies/artificial-intelligence

Question: 31
An AI practitioner has a database of animal photos. The AI practitioner wants to automatically identify and
categorize the animals in the photos without manual human effort.
Which strategy meets these requirements?

A.Object detection
B.Anomaly detection
C.Named entity recognition
D.Inpainting

Answer: A

Explanation:

The correct answer is A. Object detection because it directly addresses the requirement of automatically
identifying and categorizing objects (animals, in this case) within images.

Here's why the other options are less suitable:

B. Anomaly detection: This technique is used to identify data points that deviate significantly from the norm.
While it could potentially flag unusual animals, it wouldn't categorize common ones or identify the specific
animal type. It primarily focuses on identifying outliers, not classification.

C. Named entity recognition (NER): NER extracts named entities (like people, organizations, or locations)
from text. Since the input is images, not text, NER is not applicable.

D. Inpainting: Inpainting fills in missing parts of an image. It's a useful tool for image restoration but doesn't
classify objects within an image.

Object detection algorithms excel at identifying and localizing multiple objects within an image and assigning
a class label to each. In the context of animal photos, an object detection model could be trained to recognize
various animal species (e.g., dog, cat, bird, lion) and draw bounding boxes around each animal in the image,
along with its predicted label.

Furthermore, cloud services like Amazon Rekognition provide pre-trained object detection models that can be
readily used or fine-tuned for custom use cases. This aligns with the "AI Practitioner" context, suggesting a
practical approach. Therefore, object detection provides the necessary functionality for automatic
identification and categorization of animals in photos with minimal manual intervention.

Supporting Links:

Amazon Rekognition Object Detection: https://aws.amazon.com/rekognition/object-scene-detection/

Object Detection Tutorial: https://learnopencv.com/object-detection/
Question: 32
A company wants to create an application by using Amazon Bedrock. The company has a limited budget and
prefers flexibility without long-term commitment.
Which Amazon Bedrock pricing model meets these requirements?

A.On-Demand
B.Model customization
C.Provisioned Throughput
D.Spot Instance

Answer: A

Explanation:

The correct answer is A. On-Demand pricing for Amazon Bedrock is the most suitable option for the
company's needs. On-Demand pricing allows users to pay only for what they use, offering flexibility and
avoiding long-term commitments. This aligns perfectly with the company's desire to avoid large upfront costs
and maintain flexibility.

Option B, Model Customization, involves costs related to fine-tuning foundation models, which may not be
immediately necessary or aligned with a limited budget. Option C, Provisioned Throughput, involves
committing to a specific throughput level for a sustained period, which contradicts the company's preference
for flexibility and aversion to long-term commitment. Finally, Spot Instances (Option D) are not directly
applicable to Amazon Bedrock. Spot Instances are typically associated with EC2 and involve bidding for
unused compute capacity, which doesn't apply to the consumption model of Bedrock's AI services. The On-
Demand option offers the best balance of cost-effectiveness and flexibility.

Authoritative links:

Amazon Bedrock Pricing: https://aws.amazon.com/bedrock/pricing/

Question: 33
Which AWS service or feature can help an AI development team quickly deploy and consume a foundation model
(FM) within the team's VPC?

A.Amazon Personalize
B.Amazon SageMaker JumpStart
C.PartyRock, an Amazon Bedrock Playground
D.Amazon SageMaker endpoints

Answer: B

Explanation:

The correct answer is B. Amazon SageMaker JumpStart. Here's why:

Amazon SageMaker JumpStart provides pre-trained models, pre-built solutions, and example notebooks for a
variety of machine learning tasks, including those utilizing foundation models (FMs). It allows AI development
teams to quickly get started with FMs without needing to build everything from scratch. Crucially, SageMaker
JumpStart allows you to deploy these FMs directly within your Virtual Private Cloud (VPC), ensuring that data
and model access remain secure and compliant with your organizational policies. This is essential for many
 enterprises that need to keep their AI/ML workloads within a defined network boundary.

Amazon Personalize (A) focuses on recommendation systems, which is a specific AI use case, not a general
FM deployment tool. PartyRock (C) is a playground environment for experimenting with Amazon Bedrock and
FMs, but it's not designed for production deployment within a VPC. Amazon SageMaker endpoints (D) can be
used to deploy FMs, but SageMaker JumpStart simplifies the initial deployment process by providing pre-
configured models and deployment templates, making it much faster and easier for a team to get started. The
pre-built aspect of JumpStart significantly accelerates the time-to-market for deploying and consuming FMs
in your VPC compared to manually configuring SageMaker endpoints. In essence, JumpStart provides a
curated and simplified onboarding experience for FMs within the SageMaker ecosystem.

Further Reading:

Amazon SageMaker JumpStart Documentation

Deploying Models from SageMaker JumpStart

Question: 34
How can companies use large language models (LLMs) securely on Amazon Bedrock?

A.Design clear and specific prompts. Configure AWS Identity and Access Management (IAM) roles and policies
by using least privilege access.
B.Enable AWS Audit Manager for automatic model evaluation jobs.
C.Enable Amazon Bedrock automatic model evaluation jobs.
D.Use Amazon CloudWatch Logs to make models explainable and to monitor for bias.

Answer: A

Explanation:

The correct answer is A because security when using LLMs on Amazon Bedrock hinges on two critical
aspects: prompt engineering and access control. Designing clear and specific prompts minimizes the chances
of unintended model behavior or malicious manipulation through prompt injection. Ambiguous or overly broad
prompts increase the attack surface.

More importantly, access control via IAM roles and policies implemented using the principle of least privilege
is paramount. Least privilege means granting users and services only the permissions they absolutely need to
perform their tasks, preventing unauthorized access and data breaches. This restricts the blast radius if a
vulnerability is exploited. It confines each user and process to access only the bedrock models required for its
dedicated purpose.

While model evaluation (options B and C) is crucial for assessing performance and bias, it doesn't directly
address security vulnerabilities related to data access and model manipulation. AWS Audit Manager's focus is
on compliance auditing, not real-time model security. Similarly, while Amazon CloudWatch Logs (option D)
helps with monitoring and detecting anomalies, it doesn't prevent unauthorized access or prompt injection
attacks proactively. Explainability and bias monitoring are important aspects of responsible AI, but a strong
security posture relies on controlled access and secure prompting practices. The best approach is to reduce
the risk of misuse through prompt design combined with least privilege access.

Therefore, configuring IAM with least privilege access and designing clear prompts is the foundational
security step when leveraging LLMs on Bedrock, making option A the most pertinent response.

Further research:

AWS IAM best practices: https://docs.aws.amazon.com/IAM/latest/UserGuide/best-practices.html

 Principle of Least Privilege: https://en.wikipedia.org/wiki/Principle_of_least_privilege
Prompt Engineering Guide: https://www.promptingguide.ai/

Question: 35
A company has terabytes of data in a database that the company can use for business analysis. The company
wants to build an AI-based application that can build a SQL query from input text that employees provide. The
employees have minimal experience with technology.
Which solution meets these requirements?

A.Generative pre-trained transformers (GPT)

B.Residual neural network
C.Support vector machine
D.WaveNet

Answer: A

Explanation:

The correct answer is A, Generative Pre-trained Transformers (GPT). Here's a detailed justification:

The problem requires a solution that can translate natural language (employee input text) into SQL queries.
GPT excels at natural language processing (NLP) tasks, specifically text-to-text generation. This is because
GPT models are trained on massive datasets of text and code, enabling them to understand the relationship
between human language and structured languages like SQL. They can learn to map the intent expressed in
the input text to the corresponding SQL syntax.

The other options are not well-suited for this task. Residual neural networks (ResNets) are primarily used for
image recognition and other computer vision tasks, while Support Vector Machines (SVMs) are used for
classification and regression. Neither directly addresses the need for translating natural language to SQL.
WaveNet is designed for generating raw audio waveforms and isn't applicable to the task.

GPT's ability to perform zero-shot or few-shot learning makes it particularly attractive. It means the model
can generate SQL queries based on input prompts with minimal or no specific training on the company's data.
A fine-tuned GPT model, further trained on the company's specific database schema and business
terminology, can produce even more accurate and reliable SQL queries.

Therefore, GPT provides a ready-to-use approach for converting natural language questions into SQL, which
is essential for empowering employees with limited technical experience to access and analyze the data
effectively. Amazon offers services like Amazon Bedrock that provide access to powerful pre-trained models
including those that can perform text-to-SQL conversion.

For further research:

GPT: https://openai.com/research/gpt-3
Amazon Bedrock: https://aws.amazon.com/bedrock/
Text-to-SQL: https://arxiv.org/abs/1709.00103

Question: 36
A company built a deep learning model for object detection and deployed the model to production.
Which AI process occurs when the model analyzes a new image to identify objects?

A.Training
 B.Inference
C.Model deployment
D.Bias correction

Answer: B

Explanation:

The correct answer is B, Inference. Here's why:

Inference is the process of using a trained machine learning model to make predictions on new, unseen data.
In this scenario, the company has already built and deployed the object detection model. When the model
analyzes a new image to identify objects, it is applying its learned knowledge to classify and locate objects
within the image. This act of applying the trained model to new data to obtain predictions is precisely what
inference entails.

Training, on the other hand, is the process of creating the model itself, where it learns patterns from a dataset
of labeled examples. Model deployment is the process of making the trained model available for use. Bias
correction focuses on identifying and mitigating unfair or discriminatory outcomes from the model's
predictions or underlying data.

Since the model is already built and deployed, and the question describes it analyzing new images to identify
objects, this specifically refers to inference. The model isn't being trained, deployed, or having its bias
corrected; it's simply being used to make predictions.

Therefore, the correct answer is B.

Further reading:

Amazon SageMaker Inference: https://aws.amazon.com/sagemaker/inference/

Machine Learning Inference: https://docs.aws.amazon.com/machine-learning/latest/dg/machinelearning-
process-inference.html

Question: 37
An AI practitioner is building a model to generate images of humans in various professions. The AI practitioner
discovered that the input data is biased and that specific attributes affect the image generation and create bias in
the model.
Which technique will solve the problem?

A.Data augmentation for imbalanced classes

B.Model monitoring for class distribution
C.Retrieval Augmented Generation (RAG)
D.Watermark detection for images

Answer: A

Explanation:

The correct answer is A, Data augmentation for imbalanced classes. Here's why:

The core issue is bias stemming from imbalanced data regarding attributes within the dataset used to train
the image generation model. The model reflects the biases present in the training data.

Data augmentation addresses this directly. If, for example, the training data has fewer images of female
doctors compared to male doctors, data augmentation can artificially increase the number of images of
 female doctors. This is achieved by applying transformations to existing images of female doctors (e.g.,
rotations, zooms, slight color variations) to create new, synthetic images. By balancing the representation of
different attributes (gender, profession, race, etc.), the model is less likely to generate biased outputs. It's a
proactive approach to data imbalance.

Option B, Model monitoring for class distribution, only detects the presence of bias after the model is
deployed. It doesn't solve the underlying problem of biased training data. It can flag issues, but not prevent
them from occurring in the first place.

Option C, Retrieval Augmented Generation (RAG), is irrelevant here. RAG is typically used to improve the
factual accuracy of language models by grounding them in external knowledge sources. It's not designed to
address data imbalance or bias in image generation.

Option D, Watermark detection for images, is focused on identifying if an image has been watermarked,
possibly for copyright or authenticity purposes. It's not relevant to mitigating bias in image generation model.

Therefore, data augmentation for imbalanced classes directly tackles the root cause of the problem: the
skewed representation of attributes in the training data. By creating a more balanced dataset, the model
learns to generate images without unfairly favoring certain groups or attributes. This is a crucial step in
building fair and unbiased AI systems.

Relevant Resources:

Data Augmentation: https://www.tensorflow.org/tutorials/images/data_augmentation (TensorFlow

documentation explaining data augmentation techniques)
Fairness in Machine Learning: https://developers.google.com/machine-learning/fairness-ai (Google's guide to
understanding and mitigating bias in machine learning)

Question: 38
A company is implementing the Amazon Titan foundation model (FM) by using Amazon Bedrock. The company
needs to supplement the model by using relevant data from the company's private data sources.
Which solution will meet this requirement?

A.Use a different FM.

B.Choose a lower temperature value.
C.Create an Amazon Bedrock knowledge base.
D.Enable model invocation logging.

Answer: C

Explanation:

The correct answer is C. Create an Amazon Bedrock knowledge base. Here's why:

Amazon Bedrock knowledge bases directly address the need to augment foundation models with data from
private data sources. A knowledge base allows you to securely connect to your data repositories (like S3,
databases, etc.) and ingest the data. This ingested data is then used during inference time to ground the
foundation model's responses, making them more relevant and accurate in the context of your company's
specific information.

The Titan FM, while powerful, doesn't inherently "know" about your private data. Bedrock knowledge bases fill
this gap by providing the model with access to that data. This is a Retrieval Augmented Generation (RAG)
approach, where relevant information is retrieved from the knowledge base and provided to the model
alongside the user's prompt.
 Option A (Using a different FM) is incorrect. Switching FMs doesn't solve the problem of needing to
incorporate private data. While other FMs might have advantages, they still require a mechanism to access
and utilize company-specific knowledge.

Option B (Choosing a lower temperature value) controls the randomness of the model's output but doesn't
influence its ability to use external data sources. Lower temperature makes the output more deterministic and
less creative, but irrelevant to the data incorporation problem.

Option D (Enable model invocation logging) is for auditing and monitoring purposes. It doesn't help the model
learn from or utilize the company's private data. It records the inputs and outputs of the model, but it doesn't
enrich the model with external knowledge.

In summary, Amazon Bedrock knowledge bases are specifically designed to integrate private data with
foundation models like Amazon Titan, enabling the model to provide more informed and relevant responses.
The other options do not address this core requirement.

Further Research:

Amazon Bedrock Knowledge Bases: https://aws.amazon.com/bedrock/knowledge-bases/

Retrieval Augmented Generation (RAG): (Search for "Retrieval Augmented Generation" to find numerous
articles and research papers on the concept). While no specific AWS documentation directly defines RAG in
the context of Bedrock, knowledge bases are a prime implementation of this pattern.

Question: 39
A medical company is customizing a foundation model (FM) for diagnostic purposes. The company needs the
model to be transparent and explainable to meet regulatory requirements.
Which solution will meet these requirements?

A.Configure the security and compliance by using Amazon Inspector.

B.Generate simple metrics, reports, and examples by using Amazon SageMaker Clarify.
C.Encrypt and secure training data by using Amazon Macie.
D.Gather more data. Use Amazon Rekognition to add custom labels to the data.

Answer: B

Explanation:

The correct answer is B: Generate simple metrics, reports, and examples by using Amazon SageMaker Clarify.

Here's a detailed justification:

The question emphasizes the need for transparency and explainability of the customized foundation model
(FM). Regulatory requirements often necessitate understanding how a model arrives at its predictions,
particularly in sensitive domains like medical diagnostics.

Amazon SageMaker Clarify directly addresses this requirement. It provides tools to detect potential bias in
machine learning models and helps explain their predictions. It generates reports with feature importance
scores and examples, making the model's decision-making process more understandable. This is crucial for
demonstrating compliance with regulations that demand explainable AI.

Let's examine why the other options are not suitable:

A. Amazon Inspector: Primarily focuses on security vulnerabilities in your infrastructure and applications. It
does not provide insights into the internal workings or explainability of an AI model.
C. Amazon Macie: Is a data security and privacy service that discovers, classifies, and protects sensitive data
 stored in Amazon S3. While important for data governance, it doesn't contribute to model explainability.
D. Amazon Rekognition: Is an image and video analysis service. While potentially useful for adding labels to
data, it does not explain the decision-making process of the customized FM. More data alone doesn't
guarantee explainability; the model itself needs to be interpretable.

In summary, SageMaker Clarify is specifically designed to help understand and explain model behavior,
fulfilling the regulatory requirements for transparency and explainability, making it the most appropriate
solution in this scenario.

Supporting Links:

Amazon SageMaker Clarify: https://aws.amazon.com/sagemaker/clarify/

Explainable AI (XAI): https://aws.amazon.com/machine-learning/explainable-ai/

Question: 40
A company wants to deploy a conversational chatbot to answer customer questions. The chatbot is based on a
fine-tuned Amazon SageMaker JumpStart model. The application must comply with multiple regulatory
frameworks.
Which capabilities can the company show compliance for? (Choose two.)

A.Auto scaling inference endpoints

B.Threat detection
C.Data protection
D.Cost optimization
E.Loosely coupled microservices

Answer: BC

Explanation:

The correct answer is B. Threat detection and C. Data protection. Here's a detailed justification:

B. Threat Detection: Deploying a chatbot application, especially one dealing with potentially sensitive
customer data, necessitates robust threat detection mechanisms. Compliance frameworks often mandate
security measures to identify and mitigate potential threats, such as unauthorized access, data breaches, and
malicious attacks. Utilizing services like Amazon GuardDuty or AWS Security Hub, which can monitor API
calls, network activity, and identify potential vulnerabilities, demonstrably supports compliance regarding
threat detection. SageMaker JumpStart models themselves don't inherently provide threat detection; rather,
the deployment environment must be secured.

C. Data Protection: Data protection is a cornerstone of most regulatory frameworks. A conversational chatbot
inevitably processes and potentially stores customer data. Compliance mandates implementing appropriate
data protection measures, including encryption (both in transit and at rest), access control, and adherence to
data residency requirements. Fine-tuning a SageMaker JumpStart model does not automatically guarantee
data protection. Instead, data protection relies on how the company handles the input data to the chatbot, the
data processed during conversation, and the model's outputs. Using AWS KMS (Key Management Service) for
encryption, IAM (Identity and Access Management) for granular access control, and adhering to data
residency requirements using appropriate AWS regions are crucial for demonstrating data protection
compliance.

Why the other options are less appropriate:

A. Auto scaling inference endpoints: While auto-scaling can improve performance and availability, it's
primarily related to operational efficiency, not direct regulatory compliance. Though high availability can be
 indirectly linked to certain business continuity compliance requirements, it's not a primary compliance aspect
for chatbots concerning regulated data.

D. Cost optimization: Cost optimization is an important aspect of cloud management, but it's not a core
requirement for demonstrating compliance with regulatory frameworks pertaining to data handling and
security.

E. Loosely coupled microservices: Microservices architecture offers benefits such as scalability and
maintainability, but it doesn't directly address the core requirements of regulatory compliance in the context
of a chatbot application. The compliance focus remains on data protection and threat detection, regardless of
the application's architecture.

Supporting Links:

AWS Security: https://aws.amazon.com/security/

Amazon GuardDuty: https://aws.amazon.com/guardduty/
AWS IAM: https://aws.amazon.com/iam/
AWS KMS: https://aws.amazon.com/kms/
AWS Compliance: https://aws.amazon.com/compliance/

Question: 41
A company is training a foundation model (FM). The company wants to increase the accuracy of the model up to a
specific acceptance level.
Which solution will meet these requirements?

A.Decrease the batch size.

B.Increase the epochs.
C.Decrease the epochs.
D.Increase the temperature parameter.

Answer: B

Explanation:

The correct answer is B. Increase the epochs.

Here's a detailed justification:

The goal is to improve the accuracy of a foundation model (FM) until it reaches an acceptable level. Epochs
refer to the number of complete passes through the entire training dataset during the training process.
Increasing the number of epochs allows the model to see the data more times, refining its internal parameters
and learning the underlying patterns in the data more effectively.

Think of it like studying for an exam. The more times you review the material (the more epochs), the better
your understanding and recall become, leading to higher accuracy on the exam (the FM's accuracy). Each
epoch exposes the model to the data, allowing it to adjust its weights and biases to minimize the difference
between its predictions and the actual values (loss). By increasing epochs, the model has more opportunities
to optimize its parameters and converge towards a more accurate representation of the data.

Option A, decreasing the batch size, can impact training speed and potentially generalization, but it doesn't
directly address the fundamental need for the model to learn more from the data. A smaller batch size might
help with escaping local minima, but its primary impact isn't increasing overall accuracy.

Option C, decreasing the epochs, would be counterproductive. Fewer epochs would mean less exposure to the
training data and, therefore, a less accurate model.
 Option D, increasing the temperature parameter, is relevant to generating text or outputs from a trained
model, not to the training process itself. The temperature parameter controls the randomness of the model's
output. A higher temperature increases randomness, which is generally undesirable during training as it will
reduce the model's certainty on the correct parameters.

Therefore, increasing the epochs is the most direct and effective way to improve the accuracy of a foundation
model up to a specified level of acceptance by allowing the model to learn more thoroughly from the training
data.

For further research, consider exploring these resources:

AWS documentation on machine learning: https://aws.amazon.com/machine-learning/

Understanding Epochs in Machine Learning: https://towardsdatascience.com/epoch-vs-iterations-training-
deep-neural-networks-4a603c35f1b0 (This is a general resource on epochs)

Question: 42
A company is building a large language model (LLM) question answering chatbot. The company wants to decrease
the number of actions call center employees need to take to respond to customer questions.
Which business objective should the company use to evaluate the effect of the LLM chatbot?

A.Website engagement rate

B.Average call duration
C.Corporate social responsibility
D.Regulatory compliance

Answer: B

Explanation:

The correct answer is B. Average call duration. Here's a detailed justification:

The primary goal of the LLM-powered chatbot is to reduce the workload of call center employees by
automating responses to customer inquiries. This directly translates to decreasing the time call center
employees spend on each call. Therefore, the business objective should be a metric that reflects this
efficiency gain. Average call duration is a key performance indicator (KPI) that directly measures the length of
customer service interactions.

An LLM chatbot that successfully answers customer questions will resolve issues faster, leading to shorter
call durations. If the chatbot reduces the number of actions employees need to take (as stated in the
question), the time spent on each call must go down if the chatbot works as intended. This is because agents
will not have to search for answers or perform as many steps to assist the customer.

The other options are less relevant:

A. Website engagement rate: While a chatbot might influence website engagement, it's not the primary
focus. The core objective is call center efficiency.
C. Corporate social responsibility: While CSR is important, it's not a direct metric for evaluating the chatbot's
effectiveness in reducing call center workload.
D. Regulatory compliance: The chatbot must be compliant, but compliance is not the measuring objective.

Therefore, tracking the average call duration provides a clear and quantifiable way to assess the success of
the LLM chatbot in achieving its intended purpose: to reduce call center workload. A significant decrease in
average call duration indicates that the chatbot is effectively resolving customer issues and freeing up call
center employees to handle more complex or urgent matters.
 Authoritative Links for Further Research:

Key Performance Indicators (KPIs) in Customer Service: Search on reputable business and technology
websites (e.g., Forbes, McKinsey) for articles on KPIs in customer service.
Contact Center Metrics: Research contact center industry standards for relevant metrics.
LLM Chatbots in Customer Service: Search for case studies and articles on how LLMs are being used to
improve customer service efficiency.

Question: 43
Which functionality does Amazon SageMaker Clarify provide?

A.Integrates a Retrieval Augmented Generation (RAG) workflow

B.Monitors the quality of ML models in production
C.Documents critical details about ML models
D.Identifies potential bias during data preparation

Answer: D

Explanation:

Amazon SageMaker Clarify is specifically designed to detect and mitigate potential biases in machine
learning models and data. Its primary function is to identify biases before model training, during the data
preparation stage, and after model training, by evaluating model predictions. It achieves this by calculating a
variety of fairness metrics and providing insights into the characteristics of the data and model that might
contribute to biased outcomes. This allows data scientists to address biases proactively and build fairer, more
equitable models.

Option A is incorrect because SageMaker Clarify doesn't focus on RAG workflows. RAG is related to LLMs and
how they search and retrieve information before generating their response, which is a separate aspect. Option
B is partially true, as SageMaker Model Monitor handles ongoing model quality monitoring in production.
While Clarify can play a role in the initial model evaluation prior to deployment, the ongoing monitoring is not
its core purpose. Option C describes model card documentation which while useful, is not Clarify's primary
function. Clarify's primary focus is on bias detection and explanation, not general documentation.

Therefore, identifying potential bias during data preparation, as highlighted in option D, aligns perfectly with
SageMaker Clarify's core functionality.

Amazon SageMaker Clarify DocumentationAWS Machine Learning Blog on Fairness and Explainability

Question: 44
A company is developing a new model to predict the prices of specific items. The model performed well on the
training dataset. When the company deployed the model to production, the model's performance decreased
significantly.
What should the company do to mitigate this problem?

A.Reduce the volume of data that is used in training.

B.Add hyperparameters to the model.
C.Increase the volume of data that is used in training.
D.Increase the model training time.
 Answer: C

Explanation:

The issue described points to a case of overfitting or a shift in the data distribution between training and
production, known as data drift. The model, having performed well on the training data, fails to generalize to
unseen production data.

Option C, increasing the volume of data used in training, is the most likely solution because it helps to expose
the model to a wider range of real-world scenarios and patterns. A larger and more diverse dataset can
reduce overfitting by preventing the model from memorizing the specifics of the training data and instead
learning more generalizable features. More data increases the statistical power of the training process,
resulting in a model that's better equipped to handle variations and anomalies present in the production
environment. Furthermore, if the model is encountering new features or feature combinations in production
that were absent or underrepresented in the training data, a larger, more comprehensive dataset can help
mitigate this. This helps in capturing a more accurate and complete picture of the underlying data distribution,
which, in turn, leads to improved model performance.

Option A, reducing the data volume, is counterproductive, as it would exacerbate overfitting and reduce the
model's ability to learn generalizable patterns. Option B, adding hyperparameters, might offer some
improvement but is unlikely to address the core issue of the model's lack of exposure to real-world data
variability. Option D, increasing the training time without addressing the data issue, will not solve the problem
and might even worsen overfitting if the model simply learns the training data better without generalizing.

Therefore, while hyperparameter tuning and model architecture adjustments are crucial, the primary focus
should be on expanding the dataset to enhance the model's generalization capability and ability to handle
data drift encountered in the production environment.

Relevant resource for understanding data drift and mitigation strategies:

Amazon SageMaker Model Monitor: https://docs.aws.amazon.com/sagemaker/latest/dg/model-monitor.html

(While this focuses on monitoring, it helps understand the why of needing more data and continuously
training).

Question: 45
An ecommerce company wants to build a solution to determine customer sentiments based on written customer
reviews of products.
Which AWS services meet these requirements? (Choose two.)

A.Amazon Lex
B.Amazon Comprehend
C.Amazon Polly
D.Amazon Bedrock
E.Amazon Rekognition

Answer: BD

Explanation:

The correct answer is BD: Amazon Comprehend and Amazon Bedrock. Here's a detailed justification:

Amazon Comprehend is a natural language processing (NLP) service that uses machine learning to uncover
valuable insights from text. Its key capability, sentiment analysis, directly addresses the requirement of
determining customer sentiment from written reviews. Comprehend can identify whether a review expresses
 positive, negative, or neutral sentiment, along with a confidence score, allowing the e-commerce company to
understand customer opinions about their products at scale. It requires no machine learning expertise and is a
managed service, streamlining the process. (https://aws.amazon.com/comprehend/)

Amazon Bedrock allows you to access a wide range of foundation models (FMs) from different providers using
a single API. The service provides a single endpoint that can be used to access a diverse set of foundation
models (FMs) from AI21 Labs, Anthropic, Cohere, Meta, Stability AI, and Amazon. Foundation models can be
trained on large text datasets and can be used to solve specific business cases for generating unique and
realistic content. (https://aws.amazon.com/bedrock/) Amazon Bedrock can be used to perform sentiment
analysis using one of its foundation models.

Amazon Lex (A) is a service for building conversational interfaces using voice and text. It is used to build
chatbots but does not directly address the core requirement of analyzing existing text reviews for sentiment.

Amazon Polly (C) is a text-to-speech service that converts text into lifelike speech. While useful for other
applications, it doesn't perform sentiment analysis.

Amazon Rekognition (E) is an image and video analysis service, primarily used for object and facial
recognition, and has no direct application in analyzing text-based customer reviews for sentiment.

Question: 46
A company wants to use large language models (LLMs) with Amazon Bedrock to develop a chat interface for the
company's product manuals. The manuals are stored as PDF files.
Which solution meets these requirements MOST cost-effectively?

A.Use prompt engineering to add one PDF file as context to the user prompt when the prompt is submitted to
Amazon Bedrock.
B.Use prompt engineering to add all the PDF files as context to the user prompt when the prompt is submitted
to Amazon Bedrock.
C.Use all the PDF documents to fine-tune a model with Amazon Bedrock. Use the fine-tuned model to process
user prompts.
D.Upload PDF documents to an Amazon Bedrock knowledge base. Use the knowledge base to provide context
when users submit prompts to Amazon Bedrock.

Answer: D

Explanation:

The most cost-effective solution is option D, utilizing an Amazon Bedrock knowledge base. This approach
avoids the limitations and costs associated with the other options.

Option A is impractical because adding only one PDF file at a time limits the scope of the chat interface. Users
would not have access to information spread across multiple manuals, reducing the solution's usefulness.

Option B, embedding all PDF files in the user prompt, is also inefficient and expensive. Large prompts
consume significant resources from the LLM, leading to higher processing costs. Additionally, LLMs have
context window limitations, meaning that very long prompts could be truncated or mishandled, degrading
response quality.

Option C, fine-tuning a model using all PDF documents, is the most expensive option. Fine-tuning involves
significant compute resources and time to train a new model based on the specific dataset. This is not only
costly upfront, but also adds complexity to model management and deployment. Additionally, fine-tuning
might not be the most efficient way to retrieve information, as the LLM would need to generate the answer
based on the learned data, rather than directly retrieve relevant sections.
 Option D, leveraging a knowledge base, provides a cost-effective and efficient solution. A knowledge base
allows the LLM to access relevant information from the PDF documents on demand, instead of relying on the
full document to be included in the prompt. This reduces prompt size, lowers processing costs, and overcomes
context window limitations. Furthermore, Bedrock knowledge bases use techniques like embeddings and
vector search to quickly identify the most relevant passages within the documents, enabling more accurate
and faster responses. This approach provides a balance between cost, performance, and accuracy, making it
the optimal choice for this scenario.

Amazon Bedrock Knowledge BasesAmazon Bedrock pricing

Question: 47
A social media company wants to use a large language model (LLM) for content moderation. The company wants
to evaluate the LLM outputs for bias and potential discrimination against specific groups or individuals.
Which data source should the company use to evaluate the LLM outputs with the LEAST administrative effort?

A.User-generated content
B.Moderation logs
C.Content moderation guidelines
D.Benchmark datasets

Answer: D

Explanation:

Here's a detailed justification for why benchmark datasets are the best choice for evaluating LLM outputs for
bias and discrimination in content moderation, with minimal administrative effort:

Benchmark datasets, specifically designed for bias detection, offer a pre-existing, standardized, and readily
available resource. These datasets are meticulously curated to include diverse examples that highlight
potential biases related to gender, race, religion, and other protected attributes. Compared to user-generated
content (A) which is unstructured, potentially offensive, and requires extensive pre-processing and
annotation, benchmark datasets offer a controlled and ethical environment. Moderation logs (B), while useful
for understanding past moderation decisions, don't inherently provide a systematic way to identify new or
subtle biases in LLM output. Content moderation guidelines (C) define the expected behavior but don't
provide the data required to test if the LLM adheres to those guidelines.

Using benchmark datasets allows the company to directly compare the LLM's output against known
problematic scenarios and quantify the presence of bias using established metrics. This approach reduces the
administrative overhead of data collection, annotation, and bias identification because these tasks are already
completed. Furthermore, utilizing established benchmarks aligns with best practices in responsible AI
development. These datasets are often open-source or publicly available, minimizing cost and administrative
burden. The pre-defined format of benchmark datasets also allows for automated evaluation pipelines,
further reducing manual effort. By leveraging these datasets, the social media company can proactively
address bias concerns before deploying the LLM into production, ensuring fairer and more equitable content
moderation.

Here are some links for further research:

AI Fairness 360: https://aif360.mybluemix.net/ (Provides datasets and metrics for bias detection.)
Papers with Code - Bias in NLP: https://paperswithcode.com/task/bias-in-nlp (Lists relevant papers and
resources.)
TensorFlow Data Validation (TFDV): https://www.tensorflow.org/tfx/data_validation/get_started (Helps
identify data anomalies and biases.)
Question: 48
A company wants to use a pre-trained generative AI model to generate content for its marketing campaigns. The
company needs to ensure that the generated content aligns with the company's brand voice and messaging
requirements.
Which solution meets these requirements?

A.Optimize the model's architecture and hyperparameters to improve the model's overall performance.
B.Increase the model's complexity by adding more layers to the model's architecture.
C.Create effective prompts that provide clear instructions and context to guide the model's generation.
D.Select a large, diverse dataset to pre-train a new generative model.

Answer: C

Explanation:

The correct answer is C: Create effective prompts that provide clear instructions and context to guide the
model's generation.

Here's why:

Pre-trained generative AI models are powerful tools, but their output isn't inherently aligned with a specific
brand's voice or marketing strategy. The most direct and efficient way to control the generated content's
alignment with specific requirements is through prompt engineering. Prompts serve as instructions, providing
context and guiding the model to produce outputs that match the desired brand voice, style, and messaging.
By crafting precise and detailed prompts, the company can influence the model to generate content
consistent with its brand guidelines.

Option A (optimizing architecture and hyperparameters) focuses on improving the model's overall
performance but doesn't directly address the need for brand alignment. It is more concerned with aspects
such as reducing inference time or improving accuracy in general generation tasks, not specific styling.

Option B (increasing model complexity) could potentially improve the model's ability to learn, but it's a
computationally expensive and time-consuming approach that doesn't guarantee alignment with the brand
voice. The increase in complexity does not equate to the generation of the desired outputs.

Option D (pre-training a new model) is the most resource-intensive option. Training a new model from scratch
requires a massive dataset, significant computational resources, and a dedicated team of experts. This
approach is unnecessary when the goal is to leverage an existing model for a specific purpose. It's also
indirect; even with a carefully curated dataset, prompts are still needed to control the output.

Therefore, crafting effective prompts is the most practical and targeted solution for ensuring the generated
content adheres to the company's brand voice and messaging requirements when using a pre-trained
generative AI model.

For more information on prompt engineering, refer to these resources:

Prompt Engineering Guide: https://www.promptingguide.ai/

AWS Documentation on AI/ML: https://aws.amazon.com/machine-learning/

Question: 49
A loan company is building a generative AI-based solution to offer new applicants discounts based on specific
business criteria. The company wants to build and use an AI model responsibly to minimize bias that could
negatively affect some customers.
Which actions should the company take to meet these requirements? (Choose two.)

A.Detect imbalances or disparities in the data.

B.Ensure that the model runs frequently.
C.Evaluate the model's behavior so that the company can provide transparency to stakeholders.
D.Use the Recall-Oriented Understudy for Gisting Evaluation (ROUGE) technique to ensure that the model is
100% accurate.
E.Ensure that the model's inference time is within the accepted limits.

Answer: AC

Explanation:

The correct answer is A and C.

To minimize bias and use AI responsibly, the loan company must first detect imbalances or disparities in the
data (A). This involves analyzing the training dataset for skewed representation of certain demographics or
features that might lead to discriminatory outcomes. If the model is trained on biased data, it will likely
perpetuate and amplify those biases in its predictions. Identifying and addressing these imbalances is a
crucial step in building a fair AI model.

Further, the company should evaluate the model's behavior to provide transparency to stakeholders (C).
Model evaluation, including fairness metrics and explainability techniques, is essential to understand how the
model makes decisions and to identify potential biases in its predictions. Transparency involves documenting
the model's performance on various subgroups and communicating this information to stakeholders. This
fosters trust and allows for ongoing monitoring and refinement of the model to mitigate any negative impacts
on specific customer groups. The stakeholders will be better informed of the fairness and potential biases of
the model which is in line with responsible AI.

Option B is incorrect because the frequency of model execution has little bearing on bias or fairness. Option D
is incorrect because ROUGE is a metric used for evaluating text summarization models, and it is not relevant
to bias detection or fairness in a loan application scenario. Even if the model achieves high ROUGE scores, it
says nothing about fairness considerations. Option E is about model performance, which is about speed and
resource consumption, but not about ethical considerations.

Here are some authoritative links for further research:

Fairness in Machine Learning: https://fairlearn.org/

Amazon AI Fairness Checklist: https://aws.amazon.com/machine-learning/fairness/
Responsible AI Microsoft: https://www.microsoft.com/en-us/ai/responsible-ai

Question: 50
A company is using an Amazon Bedrock base model to summarize documents for an internal use case. The
company trained a custom model to improve the summarization quality.
Which action must the company take to use the custom model through Amazon Bedrock?

A.Purchase Provisioned Throughput for the custom model.

B.Deploy the custom model in an Amazon SageMaker endpoint for real-time inference.
C.Register the model with the Amazon SageMaker Model Registry.
D.Grant access to the custom model in Amazon Bedrock.

Answer: D
 Explanation:

The correct answer is D, granting access to the custom model in Amazon Bedrock. Here's why:

Amazon Bedrock is a fully managed service that allows you to access foundation models (FMs) from leading
AI companies, including custom models you've trained. To use a custom model with Bedrock, you need to
explicitly grant Bedrock access to it. This involves configuring the necessary permissions to allow Bedrock to
invoke your model.

Option A is incorrect because Provisioned Throughput is a mechanism used for guaranteeing inference
capacity for certain Bedrock models, not specifically a requirement for using a custom model.

Option B is incorrect because deploying the custom model to a SageMaker endpoint is an alternative method
for serving the model but not a direct requirement for using it through Bedrock. Bedrock's strength lies in
abstracting away the direct endpoint management.

Option C is incorrect because while SageMaker Model Registry is a valuable tool for managing and versioning
models, it's not a mandatory step for integrating a custom model with Amazon Bedrock. Bedrock uses a
different mechanism (resource access) for granting usage permissions. The specific implementation details
for granting access depends on where the custom model is stored, such as S3.

Fundamentally, Bedrock needs the correct permissions to use the model, regardless of the location or
registration state. This is accomplished by granting access to the model within Bedrock's configuration
settings. This aligns with the principles of least privilege and controlled access to resources within AWS.

To summarize, utilizing custom models within Amazon Bedrock relies on proper access controls, which option
D most accurately addresses. Options A, B, and C describe functionalities that are related to model
deployment but not direct requirements for Bedrock access.

Here are some resources for more information:

Amazon Bedrock documentation: https://aws.amazon.com/bedrock/

Controlling access to Amazon Bedrock resources:
https://docs.aws.amazon.com/bedrock/latest/userguide/security-iam.html

Question: 51
A company needs to choose a model from Amazon Bedrock to use internally. The company must identify a model
that generates responses in a style that the company's employees prefer.
What should the company do to meet these requirements?

A.Evaluate the models by using built-in prompt datasets.

B.Evaluate the models by using a human workforce and custom prompt datasets.
C.Use public model leaderboards to identify the model.
D.Use the model InvocationLatency runtime metrics in Amazon CloudWatch when trying models.

Answer: B

Explanation:

The correct answer is B because the company needs to identify a model that generates responses in a style
preferred by its employees, which is a subjective requirement. Evaluating models using a human workforce
and custom prompt datasets allows the company to directly assess the stylistic preferences of the
employees. Custom prompt datasets can be tailored to reflect the types of interactions employees typically
have, ensuring the evaluation is relevant. A human workforce can then judge the generated responses based
on style, tone, and overall suitability for internal use.
 Option A, using built-in prompt datasets, may not capture the specific stylistic nuances the company requires.
Public model leaderboards (option C) often focus on objective metrics like accuracy and speed and might not
reflect stylistic preferences. Option D, using InvocationLatency metrics in CloudWatch, focuses solely on the
model's response time and doesn't address the stylistic aspect at all. A human evaluation component using
custom prompts directly addresses the subjective nature of the requirement.

By employing a human workforce and custom prompt datasets, the company can gather valuable feedback on
how well each model aligns with the desired style. This approach provides a more nuanced and accurate
assessment than relying solely on automated metrics or publicly available leaderboards. This ensures the
chosen model not only performs well but also resonates with the company's employees.

For further research on evaluating language models and using human-in-the-loop approaches, you can refer
to resources on Amazon Bedrock and general best practices for evaluating AI models:

Amazon Bedrock Documentation: https://aws.amazon.com/bedrock/ (search for evaluation methods and best
practices)
Human-in-the-Loop for Machine Learning: Research papers and articles on incorporating human feedback
into the evaluation and training of AI models.

Question: 52
A student at a university is copying content from generative AI to write essays.
Which challenge of responsible generative AI does this scenario represent?

A.Toxicity
B.Hallucinations
C.Plagiarism
D.Privacy

Answer: C

Explanation:

The correct answer is C. Plagiarism.

Plagiarism, in the context of generative AI, refers to the act of using AI-generated content without proper
attribution or citation, presenting it as one's own original work. In the given scenario, the student is directly
copying content from a generative AI model and incorporating it into essays without acknowledging the
source. This directly violates academic integrity principles and constitutes plagiarism.

While other options might have tangential relevance, they don't directly address the core issue.

Toxicity refers to the generation of offensive, biased, or harmful content by the AI model. While possible with
generative AI, the scenario doesn't explicitly mention the AI producing toxic outputs.
Hallucinations occur when the AI model generates information that is factually incorrect or nonsensical.
Although AI models can hallucinate, the primary issue here is the student's unethical use of existing content,
not the AI's factual accuracy.
Privacy concerns the protection of sensitive or personal information. The scenario doesn't involve the AI
model's ability to expose private user data.

Therefore, plagiarism is the most direct and relevant challenge of responsible generative AI illustrated by the
student's actions. It emphasizes the critical need for education and ethical guidelines surrounding AI usage in
academic settings. Students need to be taught to properly cite AI-generated content and to use AI as a tool to
aid their work, not to replace original thought and composition.
 Further reading:

UNESCO AI Ethics Recommendations: https://www.unesco.org/en/articles/recommendation-ethics-artificial-

intelligence
OECD Principles on AI: https://www.oecd.org/science/oecd-principles-on-ai-5ed8f95a-en.htm

Question: 53
A company needs to build its own large language model (LLM) based on only the company's private data. The
company is concerned about the environmental effect of the training process.
Which Amazon EC2 instance type has the LEAST environmental effect when training LLMs?

A.Amazon EC2 C series

B.Amazon EC2 G series
C.Amazon EC2 P series
D.Amazon EC2 Trn series

Answer: D

Explanation:

The correct answer is D. Amazon EC2 Trn series.

Here's why:

Amazon EC2 Trn series instances are specifically designed for high-performance deep learning training and
inference, with a focus on efficiency and minimizing the environmental impact. They are powered by AWS
Trainium chips, custom-built by AWS for deep learning workloads. Trainium is optimized for training deep
learning models and excels in performance per watt.

Energy Efficiency: The Trainium chips in Trn instances are designed for energy efficiency. This means that for
the same amount of computation, they consume less power compared to other GPU-based or CPU-based
instances. Less power consumption directly translates to a lower environmental footprint, assuming the
energy source has a carbon footprint.
Purpose-Built for Deep Learning: The Trainium architecture is tailored to deep learning workloads like LLM
training. This specialization allows it to perform the required computations more efficiently than general-
purpose instances or even GPU-based instances not specifically optimized for deep learning model training.
Reduced Carbon Footprint: By using less power and performing computations efficiently, Trn instances help
reduce the carbon footprint of the training process. Training large language models can be exceptionally
energy-intensive, so selecting an energy-efficient instance type is crucial for minimizing environmental
impact.

Now let's consider why the other options are less suitable:

Amazon EC2 C series: These are compute-optimized instances that are designed for general-purpose
compute-intensive workloads. They do not necessarily have the same level of specialized energy efficiency as
the Trn series, particularly for deep learning training.
Amazon EC2 G series: These instances are designed for graphics-intensive workloads and machine learning
inference. While they can be used for training, they are generally more focused on graphics and inference
than deep learning training optimization.
Amazon EC2 P series: These are GPU-based instances that can be used for machine learning and other
compute-intensive tasks. While useful for model training, the GPUs used are often not as energy efficient or
optimized for deep learning as the Trainium chips found in Trn instances. The power consumption of GPUs
used in P series instances can be relatively high compared to specialized accelerators like Trainium.
 In summary, the Amazon EC2 Trn series instances, with their AWS Trainium chips, are the best option for a
company aiming to minimize the environmental impact of training large language models due to their superior
energy efficiency and specialization for deep learning training.

Supporting documentation:

AWS Trainium: https://aws.amazon.com/machine-learning/trainium/

Amazon EC2 Trn1 instances: https://aws.amazon.com/ec2/instance-types/trn1/

Question: 54
A company wants to build an interactive application for children that generates new stories based on classic
stories. The company wants to use Amazon Bedrock and needs to ensure that the results and topics are
appropriate for children.
Which AWS service or feature will meet these requirements?

A.Amazon Rekognition
B.Amazon Bedrock playgrounds
C.Guardrails for Amazon Bedrock
D.Agents for Amazon Bedrock

Answer: C

Explanation:

The correct answer is C. Guardrails for Amazon Bedrock.

Guardrails for Amazon Bedrock are specifically designed to implement safeguards and content moderation
policies on large language models (LLMs) accessed through Amazon Bedrock. This is crucial for applications
targeted at children, where content must be appropriate and safe. Guardrails enable you to set up deny lists
(prohibiting specific keywords or topics), toxicity filters, and hallucination detection. By defining these safety
parameters, you can ensure the stories generated by the application align with the desired content guidelines
for children.

Amazon Rekognition (A) focuses on image and video analysis and wouldn't be suitable for moderating text
generation. Amazon Bedrock Playgrounds (B) are for experimentation and model exploration but don't
inherently provide content moderation features. Agents for Amazon Bedrock (D) allow LLMs to connect to
external data sources and perform actions but don't address content safety needs directly.

Therefore, Guardrails for Amazon Bedrock is the most appropriate choice because it provides the necessary
tools to control the output of the LLM, ensuring that the generated stories are safe and appropriate for
children. It offers granular control over content, allowing you to proactively prevent the generation of
unsuitable material.

For further research on Guardrails for Amazon Bedrock, refer to the official AWS documentation:
https://aws.amazon.com/bedrock/guardrails/

Question: 55
A company is building an application that needs to generate synthetic data that is based on existing data.
Which type of model can the company use to meet this requirement?

A.Generative adversarial network (GAN)

B.XGBoost
 C.Residual neural network
D.WaveNet

Answer: A

Explanation:

The correct answer is A, Generative Adversarial Network (GAN). GANs are specifically designed to generate
new, synthetic data that resembles an existing dataset. This makes them ideal for creating synthetic data
based on real data, which is precisely what the company needs.

Here's why: GANs consist of two neural networks, a generator and a discriminator, that compete against each
other. The generator creates synthetic data, and the discriminator tries to distinguish between real and
synthetic data. Through this adversarial process, the generator learns to produce increasingly realistic
synthetic data that fools the discriminator. This allows the creation of datasets suitable for use in scenarios
where using the original data might not be permissible due to privacy or other reasons. XGBoost is a gradient
boosting algorithm generally used for classification and regression, not synthetic data generation. Residual
neural networks are typically used for image classification or similar tasks, not synthetic data. WaveNet is
designed for generating audio, not generic synthetic data based on any given input dataset.

The company's requirement is to generate synthetic data based on existing data. GANs excel at this task
because they learn the underlying distribution of the real data and generate new samples from that
distribution. XGBoost, Residual neural networks, and WaveNet do not generate synthetic data in this manner.
Therefore, GANs are the most suitable choice for the company.

For more information on GANs:

Generative Adversarial Networks: https://developers.google.com/machine-learning/gan

GANs - Stanford CS230: https://cs230.stanford.edu/section/generative-adversarial-networks-gans/

Question: 56
A digital devices company wants to predict customer demand for memory hardware. The company does not have
coding experience or knowledge of ML algorithms and needs to develop a data-driven predictive model. The
company needs to perform analysis on internal data and external data.
Which solution will meet these requirements?

A.Store the data in Amazon S3. Create ML models and demand forecast predictions by using Amazon
SageMaker built-in algorithms that use the data from Amazon S3.
B.Import the data into Amazon SageMaker Data Wrangler. Create ML models and demand forecast predictions
by using SageMaker built-in algorithms.
C.Import the data into Amazon SageMaker Data Wrangler. Build ML models and demand forecast predictions
by using an Amazon Personalize Trending-Now recipe.
D.Import the data into Amazon SageMaker Canvas. Build ML models and demand forecast predictions by
selecting the values in the data from SageMaker Canvas.

Answer: D

Explanation:

The correct answer is D, utilizing Amazon SageMaker Canvas. Here's why:

SageMaker Canvas is specifically designed for business users without coding experience to build machine
learning models. It offers a visual, point-and-click interface allowing users to directly select data and features
without writing code, directly addressing the company's lack of coding expertise. Option A, while involving
SageMaker, requires knowledge of algorithms and S3 data handling, which the company lacks. Options B and
 C use SageMaker Data Wrangler, which focuses on data preparation and also involve ML algorithm selections
beyond the company's abilities. Personalize (Option C) is geared toward personalization/recommendation and
might not be optimal for general demand forecasting.

Canvas abstracts away the complexities of ML algorithms by automatically suggesting appropriate models
based on the data provided. Users can explore different features and observe their impact on model
performance through a visual interface. The company can directly import both internal and external data into
Canvas and perform the necessary analysis for demand forecasting using its intuitive features, such as
selecting columns of data relevant to the forecast. Canvas simplifies the entire ML process, from data import
to model deployment and prediction, making it the most suitable option for the given scenario, directly
adhering to the requirement for no coding experience and the need to develop predictive models based on the
data provided. It allows for the creation of models by simply selecting values in the data, aligning with the
user's technical skill level.

Amazon SageMaker Canvas Documentation

Question: 57
A company has installed a security camera. The company uses an ML model to evaluate the security camera
footage for potential thefts. The company has discovered that the model disproportionately flags people who are
members of a specific ethnic group.
Which type of bias is affecting the model output?

A.Measurement bias
B.Sampling bias
C.Observer bias
D.Confirmation bias

Answer: B

Explanation:

Here's a breakdown of why the correct answer is sampling bias and why the other options are not suitable in
the context of the question:

Correct Answer: B. Sampling bias

Sampling bias occurs when the data used to train a machine learning model does not accurately represent the
overall population that the model will be applied to. In this case, the model disproportionately flags people of
a specific ethnic group as potential thieves. This strongly suggests that the training dataset used to build the
ML model contained an imbalance. For example, perhaps the dataset contained a significantly higher
percentage of individuals from that ethnic group who were labeled as thieves (even if they weren't). Or, the
data simply had a disproportionate representation of that ethnic group in various scenarios, leading the model
to incorrectly associate certain features or patterns with that group and a higher likelihood of theft.
Therefore, the sample of data used to train the model did not fairly represent the diversity of the actual
population it's being used to monitor, leading to the biased output. If the model was trained on a dataset that
showed a particular demographic committing theft more often, the model would learn that correlation,
regardless of whether that correlation exists in the broader, real-world population.

Why the other options are incorrect:

A. Measurement bias: Measurement bias arises when the data itself is systematically inaccurate or flawed in
a way that skews the results. For instance, if the camera quality was consistently worse when recording
individuals of a certain skin tone, leading to inaccurate feature extraction. However, the problem described
 focuses on the data used to train the model, rather than inherent flaws in the captured footage, making
sampling bias more likely.
C. Observer bias: Observer bias (also known as experimenter bias) happens when the researchers or
individuals collecting and labeling the data unconsciously influence the data in a way that confirms their pre-
existing beliefs or expectations. For example, if annotators were more likely to label actions by individuals of
a certain ethnic group as "suspicious" due to their own biases. While observer bias could contribute to biased
training data, the description explicitly points to the model disproportionately flagging a specific ethnic
group, suggesting a deeper, more fundamental issue with the composition of the data.
D. Confirmation bias: Confirmation bias is the tendency to interpret new evidence as confirmation of one's
existing beliefs or theories. While someone might experience confirmation bias when interpreting the model's
results, it doesn't explain why the model is systematically producing biased outputs in the first place.

Relevant Cloud Computing Concepts:

Data Governance: Proper data governance practices are essential for mitigating bias in machine learning
models. This includes ensuring data quality, diversity, and representativeness in training datasets.
Model Monitoring: Continuous monitoring of model performance is crucial for detecting and addressing bias.
This involves tracking metrics for different demographic groups and investigating any discrepancies.
Fairness Metrics: Various fairness metrics can be used to quantify and assess bias in machine learning
models. These metrics help ensure that the model is not unfairly discriminating against any particular group.

Authoritative Links:

AI Fairness 360: https://aif360.mybluemix.net/ - An open-source toolkit developed by IBM for detecting and
mitigating bias in machine learning models.
Google's Responsible AI Practices: https://ai.google/responsibility/ - Provides guidance on building and
deploying AI systems responsibly, including addressing fairness and bias.
AWS AI & ML Responsible AI: https://aws.amazon.com/machine-learning/responsible-ai/ - Provides tools and
resources for responsible AI development on the AWS platform.

Question: 58
A company is building a customer service chatbot. The company wants the chatbot to improve its responses by
learning from past interactions and online resources.
Which AI learning strategy provides this self-improvement capability?

A.Supervised learning with a manually curated dataset of good responses and bad responses
B.Reinforcement learning with rewards for positive customer feedback
C.Unsupervised learning to find clusters of similar customer inquiries
D.Supervised learning with a continuously updated FAQ database

Answer: B

Explanation:

Here's a detailed justification for why option B (Reinforcement learning with rewards for positive customer
feedback) is the most suitable AI learning strategy for the customer service chatbot scenario, along with
supporting concepts and links:

The core requirement is for the chatbot to improve its responses over time by learning from past interactions.
Reinforcement learning (RL) is specifically designed for this kind of iterative improvement through trial and
error. In RL, an agent (the chatbot) interacts with an environment (customer interactions), taking actions
(generating responses), and receiving feedback in the form of rewards (positive customer feedback) or
penalties (negative feedback or unresolved inquiries).
 Option B directly leverages this mechanism. The chatbot learns a policy—a mapping from states (customer
inquiries) to actions (responses)—that maximizes the cumulative reward. When a customer provides positive
feedback (e.g., a high satisfaction rating, successful resolution), the chatbot's actions leading to that outcome
are reinforced, making it more likely to produce similar responses in similar situations in the future.

Supervised learning (options A and D), while useful for initial training, relies on pre-labeled data. Option A's
curated dataset is static and doesn't allow the chatbot to learn from live interactions. Option D, continuously
updating the FAQ database, is helpful for knowledge updates but doesn't address the subtleties of
conversational AI and adapting to nuanced user requests. The chatbot will not be able to adapt to new user
queries.

Unsupervised learning (option C) can find clusters of similar inquiries, which can be helpful for topic modeling
and routing, but it doesn't provide a mechanism for the chatbot to learn better responses. This is beneficial in
discovering new topics for the chatbot to respond to, but it is not ideal for improving responses over time.

Reinforcement learning's capacity for adaptation and iterative improvement makes it ideally suited to the
dynamic nature of customer service.https://aws.amazon.com/machine-learning/reinforcement-
learning/https://azure.microsoft.com/en-us/resources/cloud-computing-dictionary/what-is-reinforcement-
learning

Question: 59
An AI practitioner has built a deep learning model to classify the types of materials in images. The AI practitioner
now wants to measure the model performance.

Which metric will help the AI practitioner evaluate the performance of the model?

A.Confusion matrix
B.Correlation matrix
C.R2 score
D.Mean squared error (MSE)

Answer: A

Explanation:

Here's a detailed justification for why a confusion matrix is the best metric for evaluating the performance of
a deep learning model classifying materials in images:

A confusion matrix is a powerful tool specifically designed for evaluating the performance of classification
models. It provides a breakdown of the model's predictions, showing the counts of true positives (TP), true
negatives (TN), false positives (FP), and false negatives (FN) for each class. This allows the AI practitioner to
see exactly where the model is making correct and incorrect predictions.

For a material classification problem, the confusion matrix would show how often the model correctly
identified each material (e.g., how many images of "wood" were correctly classified as "wood" - TP) and how
often it misclassified one material as another (e.g., how many images of "metal" were incorrectly classified as
"plastic" - FP). This granular view is crucial for understanding the model's strengths and weaknesses. It allows
for the calculation of other important performance metrics like precision, recall, F1-score, and accuracy for
each class individually and overall.

In contrast, a correlation matrix measures the linear relationship between variables, which isn't relevant for
evaluating the performance of a classification model. The R2 score (coefficient of determination) and mean
squared error (MSE) are metrics primarily used for regression problems, where the goal is to predict a
 continuous value, not to classify discrete categories. They are not suitable for assessing the accuracy of a
classification model attempting to identify different types of materials in images. Therefore, for this specific
classification task, the confusion matrix provides the most insightful and actionable information about the
model's performance. Analyzing the confusion matrix directly indicates what kinds of classification errors are
happening most frequently, facilitating model refinement by adjusting training data, architecture, or
hyperparameters.

Authoritative Links for further research:

Confusion Matrix: https://www.sciencedirect.com/topics/computer-science/confusion-matrix

Classification Evaluation Metrics: https://developers.google.com/machine-learning/crash-
course/classification/true-false-positive-negative

Question: 60
A company has built a chatbot that can respond to natural language questions with images. The company wants to
ensure that the chatbot does not return inappropriate or unwanted images.
Which solution will meet these requirements?

A.Implement moderation APIs.

B.Retrain the model with a general public dataset.
C.Perform model validation.
D.Automate user feedback integration.

Answer: A

Explanation:

The correct answer is A, implementing moderation APIs. Here's why:

The primary goal is to prevent the chatbot from returning inappropriate or unwanted images. Moderation APIs
are specifically designed to analyze images (and other content) and identify content that violates predefined
policies or guidelines. These APIs use machine learning to detect explicit content, violence, hate speech, and
other unwanted categories.

Option B, retraining the model with a general public dataset, is unlikely to directly address this problem and
could potentially introduce new biases or unwanted content. A general public dataset might contain diverse
data, but it's not guaranteed to exclude inappropriate images.

Option C, performing model validation, is a crucial step in model development, but it primarily focuses on
evaluating the model's performance on its intended task, such as accuracy and efficiency. It does not directly
address content moderation. Validation helps ensure the model works as expected, but it won't prevent the
model from selecting inappropriate images if those images are present in the dataset or if the model isn't
explicitly trained to avoid them.

Option D, automating user feedback integration, can provide valuable information about the chatbot's
performance and user experience. While helpful for identifying areas for improvement, it relies on users to
report inappropriate content after it has been shown, which is not a proactive solution to prevent such images
from being displayed in the first place.

Implementing moderation APIs (like Amazon Rekognition Image Moderation) offers a proactive approach by
filtering images before they are returned to the user. These APIs can be integrated into the chatbot's pipeline
to automatically screen images and flag those that violate specified criteria. This ensures inappropriate
images are never presented to users, satisfying the company's requirements.
 Therefore, using a moderation API provides the most direct and effective solution to the problem.

Relevant links for further research:

Amazon Rekognition Image Moderation: https://aws.amazon.com/rekognition/image-moderation/

AWS AI Services: https://aws.amazon.com/ai/

Question: 61
An AI practitioner is using an Amazon Bedrock base model to summarize session chats from the customer service
department. The AI practitioner wants to store invocation logs to monitor model input and output data.
Which strategy should the AI practitioner use?

A.Configure AWS CloudTrail as the logs destination for the model.

B.Enable invocation logging in Amazon Bedrock.
C.Configure AWS Audit Manager as the logs destination for the model.
D.Configure model invocation logging in Amazon EventBridge.

Answer: B

Explanation:

The correct answer is B. Enable invocation logging in Amazon Bedrock.

Amazon Bedrock natively supports invocation logging to directly capture and store input and output data for
model interactions. This functionality is built into the Bedrock service itself, providing a straightforward and
efficient way to monitor model usage and performance. By enabling invocation logging, the AI practitioner can
automatically record the data sent to and received from the base model, enabling auditing, debugging, and
analysis of model behavior.

Option A is incorrect because while AWS CloudTrail tracks API calls made to Bedrock, it doesn't capture the
detailed input and output data of the model invocations themselves. CloudTrail focuses on management
events (e.g., creating or deleting models) rather than data plane operations.

Option C is incorrect because AWS Audit Manager is used for automating compliance audits and doesn't
directly capture model invocation logs. It focuses on assessing and managing compliance across AWS
services, not specifically on monitoring model input and output.

Option D is incorrect because Amazon EventBridge is an event bus service, primarily used for building event-
driven applications. While EventBridge can be integrated with other AWS services, it is not the primary
mechanism for enabling and storing invocation logs for Amazon Bedrock models. While you could potentially
route invocation events to EventBridge after they are logged, it's an unnecessary complexity when Bedrock
natively supports invocation logging.

Therefore, enabling invocation logging directly within Amazon Bedrock is the most appropriate and
straightforward strategy for storing invocation logs and monitoring model input and output data. This allows
the AI practitioner to directly access and analyze the model's interactions without relying on external services
or complex configurations.

For more information, refer to the Amazon Bedrock documentation on invocation

logging:https://docs.aws.amazon.com/bedrock/latest/userguide/security-data-protection.html

Question: 62
A company is building an ML model to analyze archived data. The company must perform inference on large
datasets that are multiple GBs in size. The company does not need to access the model predictions immediately.
Which Amazon SageMaker inference option will meet these requirements?

A.Batch transform
B.Real-time inference
C.Serverless inference
D.Asynchronous inference

Answer: A

Explanation:

The correct answer is A. Batch transform.

Here's why:

Batch Transform is specifically designed for performing inference on large datasets where immediate
predictions are not required. It processes data in batches, making it suitable for scenarios involving archived
data analysis. The inference results are stored for later retrieval.
https://docs.aws.amazon.com/sagemaker/latest/dg/batch-transform.html
Real-time inference is for low-latency predictions, serving requests in real time. This is unsuitable for the
given scenario that does not need immediate access to the model predictions. Real-time inference involves
deploying a model to a persistent endpoint.
Serverless inference is suitable for applications with intermittent or unpredictable traffic patterns. While
serverless inference scales automatically, it is more ideal for applications needing low latency, which is not a
requirement in the described scenario.
Asynchronous inference addresses workloads with large payload sizes, long processing times or near real-
time latency needs. Although asynchronous inference provides asynchronous request processing and
buffering capabilities, it might not be as efficient as batch transform when dealing with large datasets where
high throughput is prioritized and lower latency is not a stringent need.

Since the company needs to analyze GBs of archived data without the need for immediate predictions, Batch
Transform offers the most cost-effective and efficient solution. It avoids the overhead of maintaining a
persistent endpoint (like real-time inference) while providing optimized processing for large batch datasets.
Batch Transform is ideal for data which is already stored, as in archived data.

Question: 63
Which term describes the numerical representations of real-world objects and concepts that AI and natural
language processing (NLP) models use to improve understanding of textual information?

A.Embeddings
B.Tokens
C.Models
D.Binaries

Answer: A

Explanation:

The correct answer is A, Embeddings. Here's why:

Embeddings are crucial in AI and NLP because they transform words, phrases, or even entire documents into
 numerical vectors. These vectors capture the semantic meaning and relationships between different elements
of textual data. AI/NLP models, particularly deep learning models, inherently operate on numbers. Therefore,
converting textual information into numerical representations is a prerequisite for these models to process
and learn from text.

Tokens (Option B) are individual units of text, like words or subwords, often used as the initial step in
preprocessing text. However, tokens themselves don't represent semantic meaning; they're merely the
building blocks for creating embeddings.

Models (Option C) are the trained algorithms that use the embeddings (among other data) to perform tasks
like text classification, sentiment analysis, or machine translation. They rely on embeddings to understand the
input text. The model itself isn't the numerical representation; it uses the numerical representation generated
by embeddings.

Binaries (Option D) refers to executable files or data represented in a binary format (0s and 1s). While machine
learning models are ultimately stored as binaries, this option doesn't describe the numerical representation of
real-world objects and concepts the models use. Binaries are a lower-level representation of the program
code itself, not the semantic data.

Embeddings, on the other hand, leverage techniques like Word2Vec, GloVe, or Transformer-based models
(like BERT) to create these numerical vectors. The position and direction of a vector in the embedding space
encode semantic similarity; words with similar meanings will have vectors that are closer together. These
vector representations allow the model to perform mathematical operations on the text data, enabling it to
learn complex patterns and relationships. Embeddings help AI understand nuanced meanings and context,
thereby improving performance in tasks involving textual information.

For further research, you can explore these resources:

Word Embeddings (TensorFlow): https://www.tensorflow.org/tutorials/text/word_embeddings

Understanding Word Vectors: https://www.analyticsvidhya.com/blog/2017/06/word-vectors-nlp-tutorial/
GloVe: Global Vectors for Word Representation: https://nlp.stanford.edu/projects/glove/

Question: 64
A research company implemented a chatbot by using a foundation model (FM) from Amazon Bedrock. The chatbot
searches for answers to questions from a large database of research papers.
After multiple prompt engineering attempts, the company notices that the FM is performing poorly because of the
complex scientific terms in the research papers.
How can the company improve the performance of the chatbot?

A.Use few-shot prompting to define how the FM can answer the questions.
B.Use domain adaptation fine-tuning to adapt the FM to complex scientific terms.
C.Change the FM inference parameters.
D.Clean the research paper data to remove complex scientific terms.

Answer: B

Explanation:

The best way to improve the chatbot's performance when struggling with complex scientific terms is B. Use
domain adaptation fine-tuning to adapt the FM to complex scientific terms. Here's why:

The core issue is the foundation model's (FM) lack of understanding of specialized scientific vocabulary.
Domain adaptation fine-tuning addresses this directly by training the FM on a dataset of research papers
containing those complex terms. This process adjusts the model's parameters to better recognize and
 process the specific language used in the research domain.

Here's why other options are less ideal:

A. Use few-shot prompting to define how the FM can answer the questions: Few-shot prompting can provide
examples, but it's unlikely to fully overcome the FM's fundamental lack of understanding of the complex
scientific terms. It's more suitable for guiding response style, not for fundamentally expanding the model's
vocabulary or comprehension.

C. Change the FM inference parameters: Modifying parameters like temperature or top-p might influence
response diversity or determinism, but it won't improve the model's comprehension of the scientific terms
themselves. Inference parameters primarily control the generation of the response, not the understanding of
the input.

D. Clean the research paper data to remove complex scientific terms: This defeats the purpose of the
chatbot. The chatbot is supposed to answer questions about the research papers, which inherently contain
complex scientific terms. Removing them would severely limit the chatbot's usefulness.

Fine-tuning, specifically domain adaptation, is designed for this exact scenario: improving a model's
performance on a specific domain (in this case, scientific research) by training it on data from that domain.
This allows the FM to learn the nuances and specific vocabulary of the research papers, leading to better
understanding and more accurate answers. It's a targeted approach to improve comprehension, rather than
merely tweaking response generation or simplifying the input data to an unacceptable degree. Fine-tuning
allows the chatbot to properly utilize all information within the provided research data.

For further research on domain adaptation and fine-tuning, you can refer to these resources:

Amazon Bedrock Documentation: https://aws.amazon.com/bedrock/ (Look for sections related to model

customization and fine-tuning. Specific documentation will depend on which FM in Bedrock is being used.)

Research Papers on Domain Adaptation in NLP: Search for academic papers on "domain adaptation" and
"fine-tuning" in the context of Natural Language Processing (NLP). Sites like Google Scholar
(https://scholar.google.com/) are great resources.

Question: 65
A company wants to use a large language model (LLM) on Amazon Bedrock for sentiment analysis. The company
needs the LLM to produce more consistent responses to the same input prompt.
Which adjustment to an inference parameter should the company make to meet these requirements?

A.Decrease the temperature value.

B.Increase the temperature value.
C.Decrease the length of output tokens.
D.Increase the maximum generation length.

Answer: A

Explanation:

The company seeks consistent LLM responses for sentiment analysis using Amazon Bedrock. The key to
achieving this lies in understanding the "temperature" parameter.

The temperature parameter controls the randomness of the LLM's output. A higher temperature (closer to 1)
introduces more randomness and creativity, making the model explore less probable word sequences. This
leads to more diverse but less predictable responses. Conversely, a lower temperature (closer to 0) makes the
 model more deterministic, favoring the most likely word sequences based on its training data. This results in
more consistent and predictable outputs, ideal when seeking repeatable results for tasks like sentiment
analysis.

Decreasing the temperature value forces the LLM to choose the most probable and predictable tokens. This
minimizes variability and helps ensure that the model returns similar responses to the same prompt each time.
This consistency is crucial for reliable sentiment scoring. Options B, C, and D do not contribute to increased
consistency. Increasing the temperature introduces more randomness. Adjusting the length of output tokens
impacts the length of the response, not its consistency. The maximum generation length limits the overall
length of the generated text, which is unrelated to the consistency in responses to the same input.

Therefore, the best way to enforce consistent LLM responses for sentiment analysis on Amazon Bedrock is by
decreasing the temperature value.

For further learning, you can consult the official Amazon Bedrock documentation regarding inference
parameters: https://docs.aws.amazon.com/bedrock/latest/userguide/model-parameters.html

Question: 66
A company wants to develop a large language model (LLM) application by using Amazon Bedrock and customer
data that is uploaded to Amazon S3. The company's security policy states that each team can access data for only
the team's own customers.
Which solution will meet these requirements?

A.Create an Amazon Bedrock custom service role for each team that has access to only the team's customer
data.
B.Create a custom service role that has Amazon S3 access. Ask teams to specify the customer name on each
Amazon Bedrock request.
C.Redact personal data in Amazon S3. Update the S3 bucket policy to allow team access to customer data.
D.Create one Amazon Bedrock role that has full Amazon S3 access. Create IAM roles for each team that have
access to only each team's customer folders.

Answer: A

Explanation:

The most secure and compliant solution for accessing customer data in Amazon S3 through Amazon Bedrock,
while adhering to team-based access restrictions, is to create separate custom service roles for each team.
This approach aligns with the principle of least privilege, granting each team access only to the specific
customer data they require.

Option A is superior because it implements fine-grained access control at the Amazon Bedrock service role
level. By creating a dedicated role for each team and restricting its S3 access to only the team's customer
data folders, the risk of unauthorized data access or modification is significantly minimized. This approach
ensures data isolation and supports compliance with the company's security policy.

Option B is flawed because it relies on teams to manually specify customer names in each request. This is
prone to human error and does not guarantee that teams will only access authorized data. It also introduces
an auditing and monitoring challenge.

Option C, redacting personal data, might address PII concerns, but it doesn't prevent unauthorized access to
other customer data within the S3 buckets. Updating the S3 bucket policy alone might not be sufficient to
enforce the team-based access control efficiently and securely.

Option D, granting full S3 access to one Bedrock role and relying on IAM roles for teams, creates an overly
 permissive Bedrock role, violating the principle of least privilege. A compromised Bedrock role could lead to
unauthorized access to all customer data.

Therefore, using separate Bedrock service roles with specific S3 access permissions for each team is the
most secure and controlled way to meet the requirements. This approach is consistent with AWS best
practices for security and access management. This reduces the blast radius of a potential compromise of a
service role.

https://docs.aws.amazon.com/IAM/latest/UserGuide/id_roles.htmlhttps://docs.aws.amazon.com/IAM/latest/UserGuide/b
practices.htmlhttps://aws.amazon.com/blogs/security/how-to-use-iam-to-grant-access-to-your-amazon-s3-
bucket/

Question: 67
A medical company deployed a disease detection model on Amazon Bedrock. To comply with privacy policies, the
company wants to prevent the model from including personal patient information in its responses. The company
also wants to receive notification when policy violations occur.
Which solution meets these requirements?

A.Use Amazon Macie to scan the model's output for sensitive data and set up alerts for potential violations.
B.Configure AWS CloudTrail to monitor the model's responses and create alerts for any detected personal
information.
C.Use Guardrails for Amazon Bedrock to filter content. Set up Amazon CloudWatch alarms for notification of
policy violations.
D.Implement Amazon SageMaker Model Monitor to detect data drift and receive alerts when model quality
degrades.

Answer: C

Explanation:

The correct answer is C because it directly addresses the requirements of preventing the model from
including personal patient information in its responses and providing notification when policy violations occur.

Here's a detailed breakdown:

Guardrails for Amazon Bedrock: Guardrails allow you to set up content filters based on defined policies. This
ensures that the model's output adheres to the privacy requirements by filtering out any personal patient
information before it is returned. This proactive content filtering prevents the model from violating privacy
policies in the first place. (Source: https://aws.amazon.com/bedrock/guardrails/)

Amazon CloudWatch Alarms: By integrating CloudWatch with the Guardrails, you can receive notifications
when a policy violation is detected. Guardrails log violation events, which CloudWatch alarms can monitor.
When a violation occurs, the CloudWatch alarm will trigger a notification, alerting the company that the model
has attempted to include personal information in its response. (Source:
https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails-cloudwatch.html)

Option A is incorrect because Amazon Macie is primarily designed for discovering and protecting sensitive
data at rest in Amazon S3 and other data stores, not for real-time filtering of model responses.

Option B is incorrect because while CloudTrail monitors API activity, it doesn't provide the content filtering
capabilities needed to prevent personal information from being included in model responses. It would only
record that a response was generated, not whether it contained sensitive data.

Option D is incorrect because SageMaker Model Monitor focuses on detecting data drift and model quality
 degradation. It doesn't address the specific requirement of preventing and alerting on the inclusion of
personal information in model responses.

Question: 68
A company manually reviews all submitted resumes in PDF format. As the company grows, the company expects
the volume of resumes to exceed the company's review capacity. The company needs an automated system to
convert the PDF resumes into plain text format for additional processing.
Which AWS service meets this requirement?

A.Amazon Textract
B.Amazon Personalize
C.Amazon Lex
D.Amazon Transcribe

Answer: A

Explanation:

The correct answer is A, Amazon Textract. Here's why:

Amazon Textract is a fully managed AWS service that uses machine learning to automatically extract text,
handwriting, and data from scanned documents. It goes beyond simple Optical Character Recognition (OCR)
to identify and extract data from forms and tables as well. This makes it perfectly suited for converting PDF
resumes, which often contain formatted text, tables (for work history), and even handwriting in some cases,
into plain text.

The company needs to convert PDF resumes into plain text for further processing, likely for tasks such as
parsing skills, experience, and contact information. Amazon Textract's ability to accurately extract text and
structured data from documents makes it the ideal solution for this use case.

Options B, C, and D are not appropriate for this scenario. Amazon Personalize is a service for creating
personalized recommendations. Amazon Lex is a service for building conversational interfaces using voice and
text. Amazon Transcribe is a service for converting speech to text. None of these directly address the
requirement of converting PDF documents to plain text.

Therefore, Amazon Textract directly addresses the need for automated PDF to text conversion, making it the
correct choice.https://aws.amazon.com/textract/

Question: 69
An education provider is building a question and answer application that uses a generative AI model to explain
complex concepts. The education provider wants to automatically change the style of the model response
depending on who is asking the question. The education provider will give the model the age range of the user who
has asked the question.
Which solution meets these requirements with the LEAST implementation effort?

A.Fine-tune the model by using additional training data that is representative of the various age ranges that the
application will support.
B.Add a role description to the prompt context that instructs the model of the age range that the response
should target.
C.Use chain-of-thought reasoning to deduce the correct style and complexity for a response suitable for that
user.
 D.Summarize the response text depending on the age of the user so that younger users receive shorter
responses.

Answer: B

Explanation:

The correct answer is B: Add a role description to the prompt context that instructs the model of the age
range that the response should target. This approach is the most efficient because it leverages the existing
capabilities of generative AI models to adapt their output based on the provided context. Prompt engineering
involves crafting effective prompts to guide the model's behavior. By including the age range in the prompt
(e.g., "Explain this concept to a 10-year-old" or "Explain this concept as if addressing a college student"), the
model can tailor its response accordingly.

Fine-tuning (option A) is a more involved process requiring significant data preparation and retraining of the
model. While it can be effective, it is overkill for simply adjusting the style based on age range and carries a
much higher implementation cost. Chain-of-thought reasoning (option C), while useful in certain scenarios, is
not directly applicable to adapting writing style based on age; it is more suited for complex problem-solving.
Summarizing the response (option D) after generation adds another layer of complexity and might lose
important nuances present in the original output. It's also a post-processing step, adding to the overall
latency.

Prompt engineering, specifically utilizing contextual information, allows for immediate and flexible control
over the model's output style. It avoids the need for model retraining or complex post-processing, making it
the solution with the least implementation effort and allowing for rapid iteration and experimentation with
different prompt styles. This method leverages the models pre-trained understanding of language styles
related to different audiences.

For further research on prompt engineering, refer to resources from leading AI providers such as OpenAI and
Google AI, which often publish guides and best practices for effectively using their models. Look for
resources on techniques like "zero-shot prompting" and "few-shot prompting," which are relevant to
understanding how to influence model behavior through prompt design.

Question: 70
Which strategy evaluates the accuracy of a foundation model (FM) that is used in image classification tasks?

A.Calculate the total cost of resources used by the model.

B.Measure the model's accuracy against a predefined benchmark dataset.
C.Count the number of layers in the neural network.
D.Assess the color accuracy of images processed by the model.

Answer: B

Explanation:

The correct strategy for evaluating the accuracy of a foundation model used in image classification is to
measure its performance against a predefined benchmark dataset (Option B). This approach allows for a
quantifiable assessment of how well the model classifies images.

Here's why:

Benchmark Datasets as Ground Truth: Benchmark datasets, such as ImageNet or CIFAR, provide a known set
of images with established labels. These datasets serve as the "ground truth" for evaluating the model's
 predictions. By comparing the model's predicted labels with the actual labels in the benchmark dataset, we
can determine its accuracy.

Quantifiable Metrics: Measuring against a benchmark dataset allows for the calculation of metrics like
accuracy, precision, recall, and F1-score. These metrics provide a clear, quantifiable measure of the model's
performance. For instance, accuracy represents the percentage of correctly classified images.

Standardized Evaluation: Using a common benchmark dataset enables standardized evaluation and
comparison of different models. This allows researchers and practitioners to objectively compare the
performance of different FMs and choose the most suitable one for their specific application.

Reproducibility: Evaluating models on publicly available benchmark datasets ensures reproducibility of

results. Others can use the same dataset to verify the reported performance and compare it with their own
models.

Option A is incorrect because cost, while important for deployment, doesn't directly reflect classification
accuracy.

Option C is incorrect because the number of layers doesn't dictate accuracy; a deeper model isn't necessarily
better.

Option D is incorrect because while color accuracy might be a factor in specific applications, the primary
metric for image classification is the correctness of the identified object, not the color representation.

Relevant Cloud Computing Concepts:

Model Evaluation Pipelines: In a cloud environment like AWS, model evaluation is typically integrated into a
model training and deployment pipeline. This pipeline often includes steps to automatically evaluate the
model against a benchmark dataset and generate performance metrics.

AWS Services: AWS offers services like Amazon SageMaker, which provides tools for model evaluation and
benchmarking.

Authoritative Links:

Amazon SageMaker Model Monitor: https://aws.amazon.com/sagemaker/model-monitor/

Machine Learning Model Evaluation: https://developers.google.com/machine-learning/crash-
course/classification/check-your-understanding

Question: 71
An accounting firm wants to implement a large language model (LLM) to automate document processing. The firm
must proceed responsibly to avoid potential harms.
What should the firm do when developing and deploying the LLM? (Choose two.)

A.Include fairness metrics for model evaluation.

B.Adjust the temperature parameter of the model.
C.Modify the training data to mitigate bias.
D.Avoid overfitting on the training data.
E.Apply prompt engineering techniques.

Answer: AC

Explanation:

The prompt asks for the best two actions an accounting firm should take when developing and deploying an
 LLM for document processing, focusing on responsible AI practices to mitigate potential harms.

Option A, "Include fairness metrics for model evaluation," is crucial. LLMs can inherit biases from their training
data, leading to discriminatory outcomes. Fairness metrics (e.g., demographic parity, equal opportunity) help
quantify these biases and inform mitigation strategies. This aligns with responsible AI principles of ensuring
equitable outcomes.

Option C, "Modify the training data to mitigate bias," is also essential. Training data often reflects societal
biases. Actively addressing and mitigating these biases in the training dataset is a fundamental step towards
building a fairer LLM. This can involve techniques like re-weighting data points, data augmentation to balance
underrepresented groups, or carefully curating datasets to remove biased content. Modifying the training
data is a preventative measure that directly addresses the source of potential bias in the LLM.

Option B, "Adjust the temperature parameter of the model," primarily affects the randomness of the model's
output, not fairness or bias. A higher temperature leads to more creative, less predictable responses, while a
lower temperature makes the responses more deterministic.

Option D, "Avoid overfitting on the training data," is a standard machine learning practice to ensure good
generalization performance, but it doesn't directly address the specific concern of responsible AI and
mitigating bias. While important for model performance, it's secondary to fairness considerations in this
context.

Option E, "Apply prompt engineering techniques," focuses on crafting effective prompts to elicit desired
responses from the LLM. While important for usability, prompt engineering doesn't inherently address bias or
fairness concerns in the underlying model.

Therefore, A and C are the best choices because they directly address responsible AI principles by focusing
on identifying and mitigating bias, ultimately working towards a fairer and more equitable AI system.

Further Reading:

AWS AI/ML Responsible AI: https://aws.amazon.com/machine-learning/responsible-ai/

Fairness Metrics: https://developers.google.com/machine-learning/fairness-overview

Question: 72
A company is building an ML model. The company collected new data and analyzed the data by creating a
correlation matrix, calculating statistics, and visualizing the data.
Which stage of the ML pipeline is the company currently in?

A.Data pre-processing
B.Feature engineering
C.Exploratory data analysis
D.Hyperparameter tuning

Answer: C

Explanation:

The answer is C: Exploratory Data Analysis (EDA).

The company's actions – creating a correlation matrix, calculating statistics, and visualizing the data – are all
core components of EDA. EDA is the crucial initial stage in the machine learning pipeline focused on
understanding the characteristics of the dataset. The primary goal is to uncover patterns, relationships,
anomalies, and potential issues within the data before applying any modeling techniques.
 Correlation matrices: Reveal relationships between different features. For example, identifying highly
correlated features can inform feature selection or highlight potential multicollinearity issues.
Calculating statistics: Providing summary measures such as mean, median, standard deviation, and quartiles
helps understand the distribution and central tendencies of the features. This helps uncover outliers and
assess the general shape of data distributions.
Visualizing the data: Enables the identification of trends, clusters, and outliers. Histograms, scatter plots, and
box plots are examples of visualizations used in EDA.

Data pre-processing (A) comes after EDA. EDA informs the data cleaning and transformation steps needed in
pre-processing. Feature engineering (B) also follows EDA and involves creating new features from existing
ones based on insights gained during EDA. Hyperparameter tuning (D) is specific to model training and
optimization, and occurs much later in the pipeline, once the model is chosen and trained. Therefore, based on
the actions described, the company is clearly engaged in the EDA phase of the machine learning pipeline.

Useful resources from AWS and others:

AWS Documentation on ML Pipelines: https://docs.aws.amazon.com/sagemaker/latest/dg/ml-pipeline.html

Overview of EDA: https://towardsdatascience.com/exploratory-data-analysis-8e25edb5906e
SageMaker Jumpstart (AWS provides various examples using this tool, including EDA):
https://aws.amazon.com/sagemaker/jumpstart/

Question: 73
A company has documents that are missing some words because of a database error. The company wants to build
an ML model that can suggest potential words to fill in the missing text.
Which type of model meets this requirement?

A.Topic modeling
B.Clustering models
C.Prescriptive ML models
D.BERT-based models

Answer: D

Explanation:

The correct answer is D, BERT-based models, because they are specifically designed for understanding and
generating text in context, a crucial requirement for filling in missing words in documents.

Let's break down why the other options are not suitable:

A. Topic Modeling: Topic modeling, like Latent Dirichlet Allocation (LDA), aims to discover abstract "topics"
within a collection of documents. It does not fill in missing words or understand the contextual relationships
needed for accurate word prediction. It groups words and documents based on shared themes but offers no
word-level prediction capability.

B. Clustering Models: Clustering algorithms (like K-Means) group data points based on similarity. While
useful for segmenting documents based on content, clustering does not predict missing words or understand
sentence structure. It focuses on grouping similar documents, not on linguistic relationships within individual
sentences.

C. Prescriptive ML Models: Prescriptive models recommend actions to take based on predicted outcomes.
These models are useful for decision-making processes but are not designed for natural language processing
tasks like text completion. They focus on optimization strategies, not on understanding and generating text.
 Now, let's explore why BERT is the right choice:

BERT (Bidirectional Encoder Representations from Transformers) is a transformer-based model pre-trained on

a massive corpus of text data. One of BERT's pre-training tasks is "Masked Language Modeling" (MLM), where
some words in the input are randomly masked, and the model is trained to predict the masked words based on
the surrounding context. This pre-training makes BERT highly effective at understanding the relationships
between words in a sentence and predicting missing words with high accuracy. BERT considers both the left
and right context (bidirectional), leading to a more nuanced understanding compared to unidirectional
language models. Finetuning a pre-trained BERT model on the specific document dataset of the company can
further improve its accuracy in suggesting suitable words.

In summary, BERT's architecture and pre-training objective (MLM) make it uniquely suited for the task of
filling in missing words in text, understanding the context, and suggesting words that fit grammatically
and semantically within the document.

Authoritative Links:

BERT: https://arxiv.org/abs/1810.04805
Hugging Face Transformers Library: https://huggingface.co/transformers/model_doc/bert.html

Question: 74
A company wants to display the total sales for its top-selling products across various retail locations in the past 12
months.
Which AWS solution should the company use to automate the generation of graphs?

A.Amazon Q in Amazon EC2

B.Amazon Q Developer
C.Amazon Q in Amazon QuickSight
D.Amazon Q in AWS Chatbot

Answer: C

Explanation:

The correct answer is C: Amazon Q in Amazon QuickSight. Here's why:

The problem requires automatically generating graphs of sales data across multiple locations over the past 12
months. Amazon QuickSight is a fully managed, cloud-native business intelligence (BI) service that makes it
easy to visualize data and derive actionable insights. Integrating Amazon Q, an AI-powered assistant, into
QuickSight enhances this capability by allowing users to ask natural language questions about their data and
receive immediate visual answers in the form of graphs and other visualizations.

Option A, Amazon Q in Amazon EC2, is incorrect because Amazon EC2 provides virtual servers in the cloud but
lacks built-in BI and visualization functionalities. While you could potentially build a visualization solution on
EC2, it wouldn't be as efficient or cost-effective as using QuickSight. Amazon Q in this context would
primarily assist with server-related tasks, not data visualization.

Option B, Amazon Q Developer, focuses on aiding software developers with code generation, debugging, and
optimization. It's not designed for business intelligence or data visualization.

Option D, Amazon Q in AWS Chatbot, facilitates interaction with AWS services through a chatbot interface.
While useful for operational tasks, it doesn't directly address the need for automated graph generation.

Amazon Q in QuickSight directly solves the problem by allowing business users to ask questions like "Show
me total sales for top-selling products by location in the last 12 months" and have QuickSight automatically
 generate the corresponding graphs using the underlying data. This empowers users without requiring them to
have deep technical expertise in data analysis or visualization.

Therefore, QuickSight with Amazon Q is the optimal solution for automating the generation of sales graphs as
it provides a seamless way to query data, create visualizations, and gain insights.

Further reading:

Amazon QuickSight: https://aws.amazon.com/quicksight/

Amazon Q: https://aws.amazon.com/amazon-q/

Question: 75
A company is building a chatbot to improve user experience. The company is using a large language model (LLM)
from Amazon Bedrock for intent detection. The company wants to use few-shot learning to improve intent
detection accuracy.
Which additional data does the company need to meet these requirements?

A.Pairs of chatbot responses and correct user intents

B.Pairs of user messages and correct chatbot responses
C.Pairs of user messages and correct user intents
D.Pairs of user intents and correct chatbot responses

Answer: C

Explanation:

The correct answer is C: Pairs of user messages and correct user intents. Here's why:

Few-shot learning involves training a machine learning model (like an LLM in Amazon Bedrock) to perform a
task using only a small number of examples. In the context of intent detection for a chatbot, the goal is to
accurately map user messages (the input) to the intended action or topic (the output).

To effectively utilize few-shot learning for intent detection, the model needs examples demonstrating the
relationship between user inputs (messages) and their corresponding intended intents. These examples serve
as the "few shots" that guide the LLM in understanding the desired mapping. For example, "I want to book a
flight to Seattle" (user message) should be paired with "book_flight" (user intent).

Option A, pairs of chatbot responses and correct user intents, is incorrect. This data focuses on the chatbot's
output rather than the user's input, which is crucial for intent detection.

Option B, pairs of user messages and correct chatbot responses, is also incorrect. While useful for training the
chatbot's response generation, it doesn't directly teach the model to understand the intent behind the user
message. It's more related to end-to-end dialogue modeling.

Option D, pairs of user intents and correct chatbot responses, is incorrect because it reverses the desired
mapping. The model needs to learn to predict the intent from the user message, not the other way around.

Therefore, having pairs of user messages and their corresponding intents is the most relevant and necessary
data for few-shot learning to improve intent detection accuracy using an LLM from Amazon Bedrock. This
type of data allows the model to learn from a limited number of examples how to accurately classify user
messages based on the underlying intent.

Further Reading:

Amazon Bedrock Documentation: https://aws.amazon.com/bedrock/ (for general understanding of the

 service)
Few-Shot Learning: Although AWS documentation may not explicitly detail "few-shot learning" directly
within the Bedrock context, searching for general resources on "few-shot learning for NLP" will provide a
good understanding of the technique used.

Question: 76
A company is using few-shot prompting on a base model that is hosted on Amazon Bedrock. The model currently
uses 10 examples in the prompt. The model is invoked once daily and is performing well. The company wants to
lower the monthly cost.
Which solution will meet these requirements?

A.Customize the model by using fine-tuning.

B.Decrease the number of tokens in the prompt.
C.Increase the number of tokens in the prompt.
D.Use Provisioned Throughput.

Answer: B

Explanation:

The correct answer is B: Decrease the number of tokens in the prompt. Here's why:

The goal is to lower the monthly cost associated with using the model on Amazon Bedrock. Since the model is
invoked once daily, the primary cost driver will be the amount of computation required for each invocation.
The size of the prompt directly impacts the computation cost. Larger prompts, containing more tokens,
require the model to process more data, thus increasing computation time and cost.

Few-shot prompting involves providing a few examples to guide the model. While 10 examples provide a
reasonable starting point, the company can likely achieve acceptable performance with fewer examples,
especially since the model is already performing well.

Decreasing the number of tokens in the prompt by reducing the number of examples reduces the amount of
data processed during each invocation. This translates directly into lower computation costs and, therefore,
lower monthly expenses.

Option A, fine-tuning, is generally more expensive upfront than simply adjusting the prompt. Fine-tuning
requires additional training data and computational resources to adapt the base model. While it might lead to
better performance with a smaller prompt later, it introduces significant initial costs.

Option C, increasing the number of tokens, is counterproductive as it would increase the computational
burden and cost.

Option D, using Provisioned Throughput, guarantees capacity and can be cost-effective if usage is high and
predictable. However, since the model is invoked only once daily, the cost benefits of Provisioned Throughput
are unlikely to outweigh the increased cost associated with reserving resources. In fact, for low usage
scenarios it's typically more expensive than on-demand pricing.

Therefore, minimizing the prompt size by reducing the number of examples is the most straightforward and
cost-effective way to lower monthly expenses without sacrificing performance.

Further Research:

Amazon Bedrock Pricing: For understanding the cost components of using Bedrock.
Prompt Engineering Techniques: Learn more about optimizing prompts for LLMs.
Question: 77
An AI practitioner is using a large language model (LLM) to create content for marketing campaigns. The
generated content sounds plausible and factual but is incorrect.
Which problem is the LLM having?

A.Data leakage
B.Hallucination
C.Overfitting
D.Underfitting

Answer: B

Explanation:

The correct answer is B, hallucination. Here's a detailed justification:

Large Language Models (LLMs) are trained on massive datasets and learn to generate text by predicting the
next word in a sequence. While they are powerful at mimicking human-like writing styles and can create
seemingly coherent content, they do not possess true understanding or access to a verified knowledge base.

Hallucination, in the context of LLMs, refers to the phenomenon where the model generates content that is
factually incorrect, nonsensical, or unrelated to the input prompt, while presenting it as if it were true. The
generated content might sound convincing, making it difficult to discern its validity.

In this scenario, the LLM produces "plausible and factual" content that is, in reality, incorrect. This directly
aligns with the definition of hallucination. The model is essentially making things up or providing information
that is not grounded in reality, even though it presents it convincingly.

Here's why the other options are less likely:

A. Data Leakage: Data leakage occurs when information from the training dataset inadvertently appears in
the output of the model. While this could manifest as incorrect information, it's not the primary characteristic
of the described problem. Data leakage is more about revealing specific training data, not necessarily
fabricating new information.
C. Overfitting: Overfitting happens when a model learns the training data too well, including the noise and
outliers. This results in poor generalization to new, unseen data. While overfitting can lead to inaccuracies, it
typically manifests as poor performance on novel data points similar to training data, rather than the
fabrication of entirely new, incorrect "facts."
D. Underfitting: Underfitting occurs when a model is too simple to capture the underlying patterns in the data.
This leads to poor performance on both the training and testing data. Underfitting would result in generalized
poor quality of generated content, but not specifically in the invention of plausible, yet incorrect, information.

Authoritative Links:

Towards Data Science - Hallucination in Large Language Models:

https://towardsdatascience.com/hallucination-in-large-language-models-explained-a98c6429d291
VentureBeat - AI Hallucinations: https://venturebeat.com/ai/why-ai-chatbots-lie-the-problem-of-ai-
hallucinations/

Therefore, hallucination is the most appropriate explanation for the LLM generating plausible but incorrect
content for marketing campaigns.
Question: 78
An AI practitioner trained a custom model on Amazon Bedrock by using a training dataset that contains
confidential data. The AI practitioner wants to ensure that the custom model does not generate inference
responses based on confidential data.
How should the AI practitioner prevent responses based on confidential data?

A.Delete the custom model. Remove the confidential data from the training dataset. Retrain the custom model.
B.Mask the confidential data in the inference responses by using dynamic data masking.
C.Encrypt the confidential data in the inference responses by using Amazon SageMaker.
D.Encrypt the confidential data in the custom model by using AWS Key Management Service (AWS KMS).

Answer: B

Explanation:

The correct answer is B. Mask the confidential data in the inference responses by using dynamic data
masking.

Here's a detailed justification:

The primary goal is to prevent the exposure of confidential data in the model's output, not to simply secure
the model itself or its training data after the fact (though those are important security considerations). The
model has already been trained with potentially sensitive data. Options A, C, and D don't directly address this
specific problem of preventing confidential information from appearing in the model's inference outputs.

Why B is correct: Dynamic data masking focuses on modifying the output data at runtime, i.e., when the model
is generating responses. This allows you to identify and redact or replace confidential elements in the
generated text before it's presented to the user. This ensures that even if the model has learned some
confidential data, it will not expose it to the outside world. This does not require re-training the model
(expensive) and works to protect data at output.

Why A is incorrect: While retraining with a cleaned dataset is a good practice for long-term model
improvement and data privacy, it's a time-consuming and resource-intensive approach. It also doesn't provide
immediate protection against existing risks associated with the current model. There is no guarantee that a
future model will not learn some level of confidential data based on surrounding context.

Why C is incorrect: Encryption of data in inference responses with Amazon SageMaker does not mask or
redact sensitive information; instead, it transforms the data into an unreadable format without the appropriate
decryption key. It prevents unauthorized access to the entire response but wouldn't allow specific elements
within the response (that aren't confidential) to be viewed. The original confidential information would still be
present, just encrypted. This is often not the desired outcome.

Why D is incorrect: Encrypting the entire custom model itself with KMS protects the model artifact from
unauthorized access. However, it does not prevent the model from generating confidential data in its
responses. Encryption alone doesn't modify the output of the model, it protects it. The confidential data would
still be generated.

In summary, dynamic data masking is the most direct and effective technique to prevent the existing trained
model from generating responses that contain confidential data. Other options involve more drastic measures
or simply do not solve the problem.

Supporting resources:

While AWS doesn't offer a specific "dynamic data masking" service for Bedrock, the concept is widely used in
data governance and security. The functionality of achieving dynamic data masking for Bedrock would involve
building custom logic to analyze generated text responses.
 For more about secure data handling and governance, refer to AWS's security best practices:
https://aws.amazon.com/security/
Explore AWS KMS for encryption basics: https://aws.amazon.com/kms/

Question: 79
A company has built a solution by using generative AI. The solution uses large language models (LLMs) to translate
training manuals from English into other languages. The company wants to evaluate the accuracy of the solution
by examining the text generated for the manuals.
Which model evaluation strategy meets these requirements?

A.Bilingual Evaluation Understudy (BLEU)

B.Root mean squared error (RMSE)
C.Recall-Oriented Understudy for Gisting Evaluation (ROUGE)
D.F1 score

Answer: A

Explanation:

The correct answer is A, Bilingual Evaluation Understudy (BLEU). Here's a detailed justification:

BLEU is specifically designed for evaluating the quality of machine-translated text. It works by comparing the
generated translation to one or more reference translations, calculating a score based on n-gram precision. In
essence, BLEU measures how many words and phrases in the machine-generated text are also present in the
human-created reference translations. A higher BLEU score indicates better similarity between the generated
and reference texts, thus better translation accuracy. In the context of translating training manuals, the
company needs a metric to assess how well the LLM is capturing the meaning and nuances of the original
English manuals in the target languages. BLEU effectively serves this purpose.

Option B, Root Mean Squared Error (RMSE), is used for evaluating regression models, where the goal is to
predict a numerical value. It measures the average magnitude of the error between predicted and actual
values. This is irrelevant for text translation, where the output is text, not numerical data.

Option C, Recall-Oriented Understudy for Gisting Evaluation (ROUGE), is primarily used for evaluating
summarization tasks. While it does compare generated text to reference text, it focuses on recall (how much
of the reference text is present in the generated text) rather than precision like BLEU. While ROUGE could
provide some insights, BLEU is more directly suited for translation quality assessment.

Option D, F1 score, is a metric used for evaluating classification models. It is the harmonic mean of precision
and recall, providing a balanced measure of the model's performance. Like RMSE, F1 score isn't appropriate
for evaluating text translation because the output is not a classification.

In summary, BLEU is the most appropriate metric because it directly addresses the task of evaluating the
quality of machine-generated translations by comparing them to reference translations, making it ideal for
assessing the accuracy of the company's LLM-powered training manual translation solution.

Further Research:

BLEU: https://aclanthology.org/P02-1040/ (Original BLEU paper)

Evaluating Text Output (Google): https://developers.google.com/machine-learning/crash-
course/classification/check-your-work (Although focused on classification, it provides context on different
evaluation metrics.)
Question: 80
A large retailer receives thousands of customer support inquiries about products every day. The customer support
inquiries need to be processed and responded to quickly. The company wants to implement Agents for Amazon
Bedrock.
What are the key benefits of using Amazon Bedrock agents that could help this retailer?

A.Generation of custom foundation models (FMs) to predict customer needs

B.Automation of repetitive tasks and orchestration of complex workflows
C.Automatically calling multiple foundation models (FMs) and consolidating the results
D.Selecting the foundation model (FM) based on predefined criteria and metrics

Answer: B

Explanation:

The correct answer is B. Automation of repetitive tasks and orchestration of complex workflows.

Here's why:

Agents for Amazon Bedrock are designed to automate tasks and streamline complex workflows. In the
context of customer support inquiries, this translates to agents being able to automatically categorize
inquiries, extract relevant information, query databases for product details or customer history, and even
draft responses. This automation significantly reduces the workload on human agents, allowing them to focus
on more complex or nuanced issues.

Option A, Generation of custom foundation models (FMs) to predict customer needs, is not the primary
function of Agents for Amazon Bedrock. While Bedrock allows for customization of FMs, Agents are about
orchestrating actions using these models, not building them from scratch.

Option C, Automatically calling multiple foundation models (FMs) and consolidating the results, is a capability
of Bedrock, but not specifically the key benefit provided by Agents. Agents use the models available through
Bedrock.

Option D, Selecting the foundation model (FM) based on predefined criteria and metrics, is also functionality
within Bedrock, but is not as central to the purpose of agents. The primary value of agents is in automating
complex operations, which includes choosing, invoking and managing FMs.

In summary, Agents for Amazon Bedrock directly address the retailer's need to quickly process and respond
to a high volume of customer support inquiries by automating repetitive tasks involved in understanding the
request, finding solutions, and delivering responses. This automated orchestration optimizes the workflow,
reducing response times and improving customer satisfaction.

For further research:

AWS documentation on Agents for Amazon Bedrock: https://aws.amazon.com/bedrock/agents/

Blog post on Agents for Amazon Bedrock: https://aws.amazon.com/blogs/aws/agents-for-amazon-bedrock-
build-agents-that-complete-tasks-for-you/

Question: 81
Which option is a benefit of ongoing pre-training when fine-tuning a foundation model (FM)?

A.Helps decrease the model's complexity

B.Improves model performance over time
C.Decreases the training time requirement
 D.Optimizes model inference time

Answer: B

Explanation:

Ongoing pre-training, even when fine-tuning a foundation model (FM), offers a crucial benefit: it improves
model performance over time. Foundation models are typically pre-trained on vast datasets representing
general knowledge. However, continuous pre-training on more relevant or specific data allows the model to
further adapt and refine its understanding of the target domain. This adaptation leads to better accuracy,
robustness, and generalization capabilities. While fine-tuning primarily focuses on adjusting the model's
parameters for a specific task, ongoing pre-training maintains and enhances the model's broader knowledge
base, preventing catastrophic forgetting and allowing it to continually learn new patterns and relationships
within the data. Options A, C, and D are less direct benefits. Pre-training alone doesn't necessarily decrease
model complexity (A); it might even increase it if new knowledge is significantly different. While tailored pre-
training can potentially reduce fine-tuning time (C) by bringing the model closer to the desired state, it’s not
the primary benefit. Pre-training isn't directly related to optimizing model inference time (D), which is typically
addressed through model optimization techniques like quantization or pruning after fine-tuning. Therefore,
the key advantage of continued pre-training during or after fine-tuning is the continuous enhancement of the
model's ability to understand and process information, leading to improved performance over the long term.

For further reading, explore resources on transfer learning and continual learning:

Transfer Learning: https://www.tensorflow.org/tutorials/images/transfer_learning (TensorFlow

documentation)
Continual Learning: Research papers and surveys can be found on platforms like arXiv and Google Scholar.

Question: 82
What are tokens in the context of generative AI models?

A.Tokens are the basic units of input and output that a generative AI model operates on, representing words,
subwords, or other linguistic units.
B.Tokens are the mathematical representations of words or concepts used in generative AI models.
C.Tokens are the pre-trained weights of a generative AI model that are fine-tuned for specific tasks.
D.Tokens are the specific prompts or instructions given to a generative AI model to generate output.

Answer: A

Explanation:

The provided answer, "Tokens are the basic units of input and output that a generative AI model operates on,
representing words, subwords, or other linguistic units," is indeed the most accurate description of tokens in
the context of generative AI.

Generative AI models, like those offered through Amazon SageMaker or accessed via APIs like Amazon
Bedrock (which offers models from AI21 Labs, Anthropic, Cohere, and Stability AI), don't process raw text
directly. Instead, text is broken down into smaller, manageable units called tokens. These tokens can be
individual words, parts of words (subwords), or even individual characters, depending on the tokenization
method used by the specific model.

The choice of tokenization method is crucial. Using words as tokens is simple but struggles with rare words or
variations. Subword tokenization, like Byte Pair Encoding (BPE) or WordPiece (used in BERT), addresses this
by breaking down rare words into more frequent subword units. This allows the model to handle out-of-
 vocabulary words more effectively.

These tokens are then mapped to numerical representations (embeddings) that the model can understand and
process mathematically. The model learns relationships between these numerical representations during
training. When generating text, the model predicts the next most probable token based on the preceding
tokens. These predicted tokens are then combined to form the output text. Therefore, tokens are not just
inputs, but also the fundamental building blocks of the model's output. Options B, C, and D are incorrect
because they either describe embeddings (B), pre-trained model weights (C), or prompts (D). While prompts
contain tokens, they aren't the definition of what a token is.

For further research, exploring resources on Natural Language Processing (NLP) and Transformer
architectures is beneficial:

Hugging Face Tokenizers: https://huggingface.co/docs/transformers/tokenizer_summary - Offers a

comprehensive overview of various tokenization techniques.
Stanford NLP: https://nlp.stanford.edu/ - Provides research and educational materials on NLP concepts.
"Attention is All You Need" (Vaswani et al., 2017): The seminal paper introducing the Transformer
architecture, which heavily relies on tokenization. Searchable on Google Scholar or similar platforms.

Question: 83
A company wants to assess the costs that are associated with using a large language model (LLM) to generate
inferences. The company wants to use Amazon Bedrock to build generative AI applications.
Which factor will drive the inference costs?

A.Number of tokens consumed

B.Temperature value
C.Amount of data used to train the LLM
D.Total training time

Answer: A

Explanation:

The correct answer is A, the number of tokens consumed. Inference costs in Amazon Bedrock, and generally
with large language models (LLMs), are primarily driven by the computational resources required to process a
request and generate a response. This resource consumption is directly proportional to the amount of text
being processed, both in the input prompt and the generated output. The text is broken down into tokens,
which can be words, parts of words, or even individual characters.

Amazon Bedrock charges you based on the number of input and output tokens processed by the selected
foundation model (FM). The more tokens an LLM consumes during inference, the more computation it needs
to perform, leading to higher costs. The model must process the input tokens to understand the context and
then generate output tokens, requiring significant matrix multiplication and other complex operations. Option
B, temperature value, influences the randomness and creativity of the model's output, but not directly the
computational cost. Options C and D, the amount of data used to train the LLM and total training time, are
factors in the model development cost, not inference cost. Inference is the process of using a trained model.
Model training has already occurred by the time inference happens.

Therefore, minimizing the length of prompts and generated outputs is the most effective way to control
inference costs when using LLMs in Amazon Bedrock.

For more information, refer to the Amazon Bedrock documentation on pricing:

https://aws.amazon.com/bedrock/pricing/
Question: 84
A company is using Amazon SageMaker Studio notebooks to build and train ML models. The company stores the
data in an Amazon S3 bucket. The company needs to manage the flow of data from Amazon S3 to SageMaker
Studio notebooks.
Which solution will meet this requirement?

A.Use Amazon Inspector to monitor SageMaker Studio.

B.Use Amazon Macie to monitor SageMaker Studio.
C.Configure SageMaker to use a VPC with an S3 endpoint.
D.Configure SageMaker to use S3 Glacier Deep Archive.

Answer: C

Explanation:

The correct answer is C. Configure SageMaker to use a VPC with an S3 endpoint.

Here's why:

The core requirement is to manage the data flow from Amazon S3 to SageMaker Studio notebooks. This
focuses on network access and security best practices for accessing S3 from within a SageMaker
environment.

Option A (Amazon Inspector): Amazon Inspector is a vulnerability management service that automates
security assessments of your applications and infrastructure. While important for security, it doesn't directly
manage the flow of data between S3 and SageMaker.

Option B (Amazon Macie): Amazon Macie is a fully managed data security and data privacy service that uses
machine learning and pattern matching to discover and protect sensitive data in AWS. Macie focuses on data
classification and security, not on managing the data transfer pathway itself.

Option D (S3 Glacier Deep Archive): S3 Glacier Deep Archive is a storage class optimized for infrequent data
access. While useful for long-term archival, using it doesn't help with managing the data flow to SageMaker
notebooks.

Why Option C is the best:

Configuring SageMaker to use a VPC (Virtual Private Cloud) with an S3 endpoint provides a secure and
controlled network path for SageMaker Studio notebooks to access data in S3.

1. VPC Isolation: A VPC allows you to create a private, isolated network within AWS. This limits the
attack surface of your SageMaker environment by controlling network ingress and egress.

2. S3 Endpoint (Gateway or Interface): A VPC endpoint for S3 creates a direct, private connection
between your VPC and S3, without requiring traffic to traverse the public internet. This significantly
enhances security and improves performance because it avoids internet bandwidth limitations and
potential exposure. Specifically, a gateway endpoint for S3 is free and optimized for bulk data
transfer, ideal for moving datasets to SageMaker.

3. Data Flow Management: By restricting access to S3 through the VPC and endpoint, you control the
flow of data, ensuring only authorized instances (like your SageMaker Studio notebook) can access
the S3 bucket. You can further refine access using IAM roles and bucket policies.

In summary, using a VPC with an S3 endpoint establishes a secure and manageable pathway for SageMaker
to access data in S3, directly addressing the company's needs.
 Authoritative Links:

Amazon SageMaker Security

Using Amazon S3 on VPC
Controlling access to Amazon S3 resources

Question: 85
A company has a foundation model (FM) that was customized by using Amazon Bedrock to answer customer
queries about products. The company wants to validate the model's responses to new types of queries. The
company needs to upload a new dataset that Amazon Bedrock can use for validation.
Which AWS service meets these requirements?

A.Amazon S3
B.Amazon Elastic Block Store (Amazon EBS)
C.Amazon Elastic File System (Amazon EFS)
D.AWS Snowcone

Answer: A

Explanation:

The correct answer is A. Amazon S3.

Amazon S3 (Simple Storage Service) is a highly scalable, durable, and cost-effective object storage service. It
is ideal for storing the dataset needed to validate the customized foundation model in Amazon Bedrock.
Here's why:

1. Storage for Unstructured Data: Datasets for model validation often consist of text files (e.g., CSV,
JSON), images, or other unstructured data. S3 is designed for this type of data.

2. Scalability and Durability: S3 can accommodate large datasets of any size and provides high
durability (99.999999999%) to ensure data isn't lost. This is critical for long-term storage of
datasets.

3. Integration with AWS Services: S3 seamlessly integrates with other AWS services, including
Amazon Bedrock. Bedrock can directly access datasets stored in S3 for model validation.

4. Cost-Effectiveness: S3 offers various storage classes (e.g., S3 Standard, S3 Intelligent-Tiering, S3

Glacier) that allow users to optimize storage costs based on access frequency.

5. Security: S3 provides security features such as access control lists (ACLs) and bucket policies to
control access to the dataset.

Why other options are not suitable:

B. Amazon Elastic Block Store (Amazon EBS): EBS provides block-level storage volumes for use with Amazon
EC2 instances. It's not the appropriate service for storing datasets for model validation. EBS is primarily for
persistent storage of operating systems, applications, and data accessed by EC2 instances.
C. Amazon Elastic File System (Amazon EFS): EFS provides a scalable, elastic, shared file system for use
with AWS Cloud services and on-premises resources. While it can store data, it's generally more expensive
than S3 and designed for shared file system access, which isn't the primary requirement here.
D. AWS Snowcone: AWS Snowcone is a small, rugged, and secure edge computing and data transfer device.
It's designed for collecting data in edge locations and transferring it to AWS. It's not suitable for ongoing
storage and access of validation datasets within AWS.
 In conclusion, Amazon S3's scalability, durability, cost-effectiveness, and seamless integration with Amazon
Bedrock make it the best choice for storing and accessing the dataset needed to validate the foundation
model's responses.

Authoritative Links:

Amazon S3: https://aws.amazon.com/s3/

Amazon Bedrock: https://aws.amazon.com/bedrock/

Question: 86
Which prompting attack directly exposes the configured behavior of a large language model (LLM)?

A.Prompted persona switches

B.Exploiting friendliness and trust
C.Ignoring the prompt template
D.Extracting the prompt template

Answer: D

Explanation:

The question asks which prompting attack reveals an LLM's configured behavior. The correct answer is (D)
Extracting the prompt template. Here's why:

Prompt templates are blueprints that guide an LLM's output by structuring the input and specifying desired
response styles, formatting, and constraints. They essentially define the LLM's role and behavior. Extracting
this template means uncovering the underlying instructions the LLM is operating under.

Other options are less direct. Prompted persona switches (A) involve manipulating the LLM to adopt different
roles, but this doesn't necessarily expose the configured behavior (i.e., the underlying template). Exploiting
friendliness and trust (B) relies on social engineering to elicit desired responses, rather than revealing the
LLM's structural setup. Ignoring the prompt template (C), if successful, might lead to unpredictable outputs
but doesn’t reveal what the original template was.

Extracting the template, however, directly unveils the internal structure, showcasing how the LLM is intended
to process input and generate output. This exposes the model's constraints, guidelines, and pre-defined
knowledge or biases embedded within the template. For instance, if the prompt template instructs the LLM to
always provide answers in a concise, bulleted list, extracting the template will reveal this configured behavior.
This is significant as it helps attackers understand the system's vulnerabilities and potentially bypass security
measures. Furthermore, knowing the underlying template, an attacker can craft inputs tailored to manipulate
the LLM more effectively.

For further reading, research on prompt injection and LLM security can provide deeper insights into these
vulnerabilities:

OWASP: https://owasp.org/www-project-top-ten/ (Although not specific to LLMs, it provides foundational

understanding of web application vulnerabilities)
Research papers on prompt injection and LLM attacks (search on Google Scholar or Arxiv).

Question: 87
A company wants to use Amazon Bedrock. The company needs to review which security aspects the company is
responsible for when using Amazon Bedrock.
Which security aspect will the company be responsible for?

A.Patching and updating the versions of Amazon Bedrock

B.Protecting the infrastructure that hosts Amazon Bedrock
C.Securing the company's data in transit and at rest
D.Provisioning Amazon Bedrock within the company network

Answer: C

Explanation:

The correct answer is C: Securing the company's data in transit and at rest. This aligns with the shared
responsibility model in cloud computing, especially in the context of managed services like Amazon Bedrock.
Amazon Bedrock handles the underlying infrastructure security (OS, network, and physical security),
patching, updating, and the operational maintenance of the Bedrock service itself (Answers A & B).

However, the responsibility of securing the data used with the service always rests with the customer. This
includes encrypting data both while it's being transferred to and from Bedrock (in transit) and while it's stored
(at rest), whether it's prompts, input data, or the outputs generated by the models. The customer must
implement appropriate access controls, authentication, and authorization mechanisms to protect their data
from unauthorized access. Answer D, provisioning Bedrock within the company network, isn't applicable, as
Bedrock is accessed via AWS APIs and doesn't reside within a company's network. AWS is responsible for the
security of the cloud, while the customer is responsible for security in the cloud. Specifically concerning data,
the customer is always responsible for its classification, protection, and lifecycle management.

Further reading on the AWS Shared Responsibility Model can be found here:

AWS Shared Responsibility Model

Amazon Bedrock Security

Question: 88
A social media company wants to use a large language model (LLM) to summarize messages. The company has
chosen a few LLMs that are available on Amazon SageMaker JumpStart. The company wants to compare the
generated output toxicity of these models.

Which strategy gives the company the ability to evaluate the LLMs with the LEAST operational overhead?

A.Crowd-sourced evaluation
B.Automatic model evaluation
C.Model evaluation with human workers
D.Reinforcement learning from human feedback (RLHF)

Answer: B

Explanation:

The correct answer is B. Automatic model evaluation. Here's a detailed justification:

The question emphasizes minimizing operational overhead while evaluating LLM toxicity. Automatic model
evaluation directly addresses this requirement. Services like Amazon SageMaker Model Monitor and other
automated tools provide metrics on model performance, including toxicity, without significant manual
intervention. These tools can be configured to automatically analyze model outputs and generate reports
highlighting potential issues like toxic content.
 Crowd-sourced evaluation (A) and model evaluation with human workers (C) involve significant manual effort.
They require setting up platforms for human reviewers, providing guidelines, and managing the review
process, which adds to the operational overhead.

Reinforcement learning from human feedback (RLHF) (D) is a training technique used to align LLMs with
human preferences, but it primarily focuses on improving model behavior during training, not directly
evaluating toxicity after deployment. While RLHF can reduce toxicity over time, it isn't a rapid evaluation
strategy. It also involves complex implementation and continuous retraining, significantly increasing
overhead.

Automatic model evaluation tools can be integrated into the LLM deployment pipeline on SageMaker,
enabling continuous monitoring of toxicity levels with minimal manual effort. This strategy allows the social
media company to quickly compare toxicity metrics across the chosen LLMs available on SageMaker
JumpStart. Automated analysis provides quantitative data, enabling data-driven decisions on model selection.
The automated evaluation minimizes the labor cost as well, further reducing operational expenses.

Therefore, automatic model evaluation provides the most efficient and cost-effective way to compare the
generated output toxicity of the LLMs with the least operational overhead.

Supporting documentation:

Amazon SageMaker Model Monitor: https://docs.aws.amazon.com/sagemaker/latest/dg/model-monitor.html

SageMaker JumpStart: https://aws.amazon.com/sagemaker/jumpstart/

Question: 89
A company is testing the security of a foundation model (FM). During testing, the company wants to get around the
safety features and make harmful content.

Which security technique is this an example of?

A.Fuzzing training data to find vulnerabilities

B.Denial of service (DoS)
C.Penetration testing with authorization
D.Jailbreak

Answer: D

Explanation:

The correct answer is D. Jailbreak.

Jailbreaking, in the context of foundation models, refers to techniques designed to circumvent the model's
safety mechanisms and generate outputs that violate its intended ethical guidelines or safety policies. The
scenario describes an attempt to bypass the built-in safety features to produce harmful content, which aligns
directly with the goal of jailbreaking attacks. It's a form of adversarial attack that focuses on prompt
engineering to manipulate the model's behavior.

Option A, fuzzing training data, involves introducing random or unexpected data variations into the model's
training dataset. While fuzzing can uncover vulnerabilities, it is not directly intended to bypass safety features
at inference time. It's more about finding weaknesses during the training phase.

Option B, denial of service (DoS), is an attack aimed at making a service unavailable to legitimate users,
typically by overwhelming it with traffic. This is an infrastructure level attack and not related to the FM's
inherent safety features.
 Option C, penetration testing with authorization, is a security assessment where authorized testers simulate
attacks to identify vulnerabilities. While it is related to security, it is a broader process and doesn't specifically
denote attempts to circumvent safety features and generate harmful content like jailbreaking does. A
penetration test may include jailbreaking but it is not necessarily the central activity.

Therefore, the best description for the scenario is jailbreaking.For further research on jailbreaking FMs and
adversarial attacks, refer to resources such as:

OWASP: https://owasp.org/ (While OWASP doesn't have a specific page on FM jailbreaking yet, they have
resources on prompt injection and similar attacks that are relevant.)
Research papers on adversarial attacks and prompt engineering: Search on Google Scholar or ArXiv for
papers related to "adversarial attacks on large language models" or "prompt injection."

Question: 90
A company needs to use Amazon SageMaker for model training and inference. The company must comply with
regulatory requirements to run SageMaker jobs in an isolated environment without internet access.

Which solution will meet these requirements?

A.Run SageMaker training and inference by using SageMaker Experiments.

B.Run SageMaker training and Inference by using network Isolation.
C.Encrypt the data at rest by using encryption for SageMaker geospatial capabilities.
D.Associate appropriate AWS Identity and Access Management (IAM) roles with the SageMaker jobs.

Answer: B

Explanation:

The correct answer is B. Run SageMaker training and Inference by using network Isolation.

Here's why: The requirement is to run SageMaker jobs in a completely isolated environment without internet
access to comply with regulatory requirements. This necessitates network isolation.

Network Isolation (Option B) directly addresses the isolation requirement. SageMaker offers several ways to
achieve network isolation, primarily through the use of VPCs (Virtual Private Clouds). You can configure
SageMaker notebooks, training jobs, inference endpoints, and processing jobs within your VPC, restricting all
inbound and outbound internet traffic. This meets the requirement of running jobs without internet access.
VPC configurations would include setting up VPC endpoints for accessing other AWS services (like S3 for
data storage) without traversing the public internet. This solution is designed for situations needing enhanced
security or compliance.

SageMaker Experiments (Option A) is used for tracking, organizing, and comparing different machine
learning experiments. While valuable for managing the model development lifecycle, it doesn't inherently
provide network isolation.

Encryption for SageMaker geospatial capabilities (Option C) focuses on securing data at rest for geospatial
data. While encryption is crucial for data security, it doesn't directly isolate the network, and it is relevant to a
specific subset of data types.

IAM roles (Option D) are vital for granting permissions to SageMaker jobs to access AWS resources, they do
not provide network isolation. IAM governs what resources SageMaker can access, not how it accesses them.
Even with restrictive IAM roles, the job could still theoretically have network access if the underlying network
isn't isolated.
 Therefore, only option B directly satisfies the requirement of running SageMaker jobs in an environment
isolated from the internet.

Supporting Links:

Secure Amazon SageMaker resources by using VPC:

https://docs.aws.amazon.com/sagemaker/latest/dg/security-vpc.html
Amazon VPC Endpoints: https://docs.aws.amazon.com/vpc/latest/userguide/vpc-endpoints.html

Question: 91
An ML research team develops custom ML models. The model artifacts are shared with other teams for integration
into products and services. The ML team retains the model training code and data. The ML team wants to build a
mechanism that the ML team can use to audit models.

Which solution should the ML team use when publishing the custom ML models?

A.Create documents with the relevant information. Store the documents in Amazon S3.
B.Use AWS AI Service Cards for transparency and understanding models.
C.Create Amazon SageMaker Model Cards with intended uses and training and inference details.
D.Create model training scripts. Commit the model training scripts to a Git repository.

Answer: C

Explanation:

The correct answer is C: Create Amazon SageMaker Model Cards with intended uses and training and
inference details.

Here's a detailed justification:

The ML team requires a robust auditing mechanism for the custom ML models that they develop and share.
This mechanism must track the model's intended use, training data details, inference configuration, and other
relevant information throughout the model's lifecycle.

Option C directly addresses this requirement by leveraging Amazon SageMaker Model Cards. Model Cards
are designed specifically for documenting and auditing ML models. They provide a structured way to record
crucial details such as model purpose, training data provenance, evaluation metrics, intended use cases,
limitations, and versioning. This comprehensive documentation enables the ML team to effectively track and
understand the model's characteristics, making auditing significantly more manageable.

Option A, creating documents and storing them in Amazon S3, is a rudimentary solution. While it provides
storage, it lacks the structured and standardized format crucial for effective auditing. Searching and
consolidating information across disparate documents in S3 is inefficient and error-prone.

Option B, using AWS AI Service Cards, is relevant for transparency and understanding. However, AI Service
Cards primarily focus on the general capabilities of AWS AI services, not custom-built models. They might not
offer the level of detail the ML team needs to fully audit the specific models developed.

Option D, committing training scripts to a Git repository, is a good practice for version control of the code, but
it doesn't capture essential model metadata like intended use, training data details, and evaluation metrics.
It's more focused on code provenance rather than comprehensive model auditing.

SageMaker Model Cards offer a centralized, structured, and detailed record for each model, enabling
efficient auditing and ensuring models are used responsibly and within their intended parameters. The model
card captures a single source of truth for the model and its characteristics, making auditing much easier.
 Therefore, Amazon SageMaker Model Cards provide the most comprehensive and suitable solution for the ML
team's auditing needs, ensuring accountability, transparency, and responsible use of their custom ML models.

Authoritative links:

Amazon SageMaker Model Cards: https://aws.amazon.com/sagemaker/model-cards/

AWS AI Service Cards: https://aws.amazon.com/machine-learning/responsible-ai/

Question: 92
A software company builds tools for customers. The company wants to use AI to increase software development
productivity.

Which solution will meet these requirements?

A.Use a binary classification model to generate code reviews.

B.Install code recommendation software in the company's developer tools.
C.Install a code forecasting tool to predict potential code issues.
D.Use a natural language processing (NLP) tool to generate code.

Answer: B

Explanation:

The correct answer is B. Install code recommendation software in the company's developer tools.

Here's a detailed justification:

The core goal is to increase software development productivity using AI. Code recommendation software
directly addresses this by suggesting code snippets, libraries, and patterns to developers as they write code.
This helps automate mundane coding tasks, reduces errors, and accelerates the development process. These
systems learn from existing codebases and developer habits to provide contextually relevant suggestions.

Option A, using a binary classification model for code reviews, can be helpful but is more focused on quality
assurance after the code is written, not directly improving development speed. While valuable, it doesn't
provide real-time assistance during coding like a recommendation system.

Option C, a code forecasting tool, aims to predict potential issues. While this proactive approach could save
time in the long run, it doesn't actively assist the developer during the coding phase, hence provides less
immediate productivity gains.

Option D, using NLP to generate code, is a very complex and often unreliable approach for general-purpose
code generation. While NLP is advancing, current tools are not robust enough to generate entire code
segments with high accuracy and maintainability for complex applications. It's more suited for specific, well-
defined tasks.

Code recommendation systems, on the other hand, are readily available and proven tools for enhancing
developer productivity. They fit directly into the existing development workflow, integrating with IDEs and
other developer tools. Examples of such tools are integrated in IDEs like VS Code through plugins such as
GitHub Copilot. This is the most direct way to impact development speed, accuracy and efficiency.

For more information on AI-powered code assistants and their benefits, consider researching tools such as
GitHub Copilot, Tabnine, and AWS CodeWhisperer. These tools showcase the practical application of AI in
improving software development productivity.

Supporting Resources:
 GitHub Copilot: https://github.com/features/copilot
Tabnine: https://www.tabnine.com/
AWS CodeWhisperer: https://aws.amazon.com/codewhisperer/

Question: 93
A retail store wants to predict the demand for a specific product for the next few weeks by using the Amazon
SageMaker DeepAR forecasting algorithm.

Which type of data will meet this requirement?

A.Text data
B.Image data
C.Time series data
D.Binary data

Answer: C

Explanation:

The correct answer is C, Time series data, because Amazon SageMaker DeepAR is specifically designed for
time series forecasting. DeepAR is a supervised learning algorithm that uses recurrent neural networks
(RNNs) to predict future values based on past sequential data. Time series data represents a sequence of data
points indexed in time order.

Here's why the other options are incorrect:

A. Text data: Text data is unstructured and not inherently suitable for forecasting future numerical values
without significant preprocessing and transformation. While you could extract signals from text data (e.g.,
sentiment analysis impacting sales), it's not the primary input for a DeepAR model used for demand
forecasting.
B. Image data: Image data is irrelevant for demand forecasting based on historical trends.
D. Binary data: Binary data represents two states (0 or 1) and is not suitable for representing continuous
values or trends over time, which are essential for demand forecasting.

For DeepAR to effectively predict demand, the retail store needs to provide historical sales data for the
product, indexed by date or time. Each data point would represent the sales volume or demand for the
product at a specific point in time. The model learns patterns and seasonality from this historical time series
data to forecast future demand. For instance, it can identify trends like increased sales during holidays or
specific seasons. The richer the historical time series, the more accurate the prediction typically. DeepAR also
handles related time series (e.g., price, promotions), which can improve forecast accuracy.

Relevant Links:

Amazon SageMaker DeepAR Documentation:

https://docs.aws.amazon.com/sagemaker/latest/dg/deepar.html
DeepAR: Probabilistic Forecasting with Autoregressive Recurrent Networks:
https://arxiv.org/abs/1704.04110 (Original Research Paper)

Question: 94
A large retail bank wants to develop an ML system to help the risk management team decide on loan allocations
for different demographics.
What must the bank do to develop an unbiased ML model?

A.Reduce the size of the training dataset.

B.Ensure that the ML model predictions are consistent with historical results.
C.Create a different ML model for each demographic group.
D.Measure class imbalance on the training dataset. Adapt the training process accordingly.

Answer: D

Explanation:

Here's a detailed justification for why option D is the most appropriate choice for developing an unbiased ML
model for loan allocation, considering the context of risk management in a large retail bank:

Detailed Justification:

Option D, "Measure class imbalance on the training dataset. Adapt the training process accordingly," is the
most crucial first step in mitigating bias in this ML model. Class imbalance refers to situations where certain
demographic groups (classes) are underrepresented in the training data. This skew can lead the model to
unfairly favor the majority group, perpetuating existing biases in loan allocations. Measuring this imbalance
allows the bank to quantify the problem.

Once measured, adaptation strategies can be employed. These include:

Resampling techniques: Oversampling the minority class (e.g., SMOTE) or undersampling the majority class
to create a more balanced dataset.
Cost-sensitive learning: Assigning higher penalties for misclassifying the underrepresented group.
Using evaluation metrics appropriate for imbalanced datasets: Focusing on metrics like precision, recall, F1-
score, and AUC that provide a more nuanced view of model performance than simple accuracy, especially
when dealing with class imbalance.

Reducing the dataset size (option A) can worsen the problem, particularly if the already underrepresented
groups are further marginalized. Ensuring consistency with historical results (option B) is dangerous, as
historical results may already reflect existing biases, and replicating them would simply perpetuate
unfairness. Creating separate models for each demographic group (option C) raises ethical concerns
regarding potential discrimination and is often legally problematic.

Addressing class imbalance doesn't guarantee a completely unbiased model, but it's a fundamental step
toward fairer and more equitable loan allocations. Further steps include careful feature selection to avoid
using proxies for sensitive attributes (like zip code as a proxy for race) and continuous monitoring for bias in
the deployed model. By carefully measuring and addressing class imbalance, the bank can build a more
responsible and ethical ML system for loan allocation.

Authoritative Links for Further Research:

1. Amazon SageMaker Clarify: This AWS service can detect potential bias in ML
models.https://aws.amazon.com/sagemaker/clarify/
2. Fairness in Machine Learning (Google AI): Provides resources and best practices for building fair ML
systems.https://ai.google/responsibility/ai-principles/
3. Dealing with Imbalanced Data: Explains various techniques for addressing class
imbalance.https://developers.google.com/machine-learning/data-prep/transform/class-imbalance
4. Bias Mitigation Techniques by Microsoft: A broad overview of different mitigation approaches for
bias in ML systems.https://www.microsoft.com/en-us/research/uploads/prod/2019/03/Bias-
Mitigation-Techniques.pdf
Question: 95
Which prompting technique can protect against prompt injection attacks?

A.Adversarial prompting
B.Zero-shot prompting
C.Least-to-most prompting
D.Chain-of-thought prompting

Answer: A

Explanation:

The correct answer is A. Adversarial prompting helps mitigate prompt injection attacks by specifically
crafting prompts designed to expose vulnerabilities in a language model's defenses. By actively testing the
model's response to potentially malicious inputs, developers can identify weaknesses and implement
safeguards. The goal is to trick the AI into bypassing its intended programming and revealing sensitive
information or performing unauthorized actions. This proactive approach allows for the detection and
patching of prompt injection vulnerabilities before malicious actors can exploit them. While zero-shot, least-
to-most, and chain-of-thought prompting are valuable techniques for improving the model's reasoning and
generation capabilities, they do not directly address the security risks associated with prompt injection.
Adversarial prompting, in contrast, directly simulates potential attacks, thereby fortifying the AI system
against such threats. Think of it as a "stress test" for the AI's security. By understanding how the AI reacts to
adversarial inputs, developers can build more resilient AI applications.

Reference links:

OWASP Prompt Injection: https://owasp.org/www-project-top-10-for-large-language-model-applications/

Research paper on adversarial attacks on LLMs: (Search on Google Scholar for "adversarial attacks large
language models" for up-to-date research)

Question: 96
A company has fine-tuned a large language model (LLM) to answer questions for a help desk. The company wants
to determine if the fine-tuning has enhanced the model's accuracy.

Which metric should the company use for the evaluation?

A.Precision
B.Time to first token
C.F1 score
D.Word error rate

Answer: C

Explanation:

The question asks about evaluating the accuracy enhancement of a fine-tuned LLM used for help desk
question answering. The optimal metric should reflect both precision and recall in providing correct answers.

F1 Score (Option C): The F1 score is the harmonic mean of precision and recall. Precision measures how many
of the answers the LLM provides are correct, while recall measures how many of the correct answers the LLM
does provide. A high F1 score signifies a good balance between precision and recall, indicating that the fine-
tuning has improved the LLM's ability to provide accurate and complete answers to help desk questions. If the
 model is too focused on precision, it might avoid answering many questions, hurting recall. If it's too focused
on recall, it might answer many questions inaccurately, hurting precision. F1 score elegantly balances these
concerns.

Precision (Option A): Precision alone only tells us how often the answers are correct when the model provides
them. It doesn't tell us if the model is missing many correct answers. For example, if the model only answers a
few questions but answers them correctly, precision would be high, but it wouldn't be a good help desk
assistant.

Time to First Token (Option B): This metric measures the latency before the LLM starts generating the
answer. While important for user experience, it doesn't assess the accuracy of the answer itself. It's a
performance metric, not an accuracy metric.

Word Error Rate (WER) (Option D): WER is primarily used to evaluate speech recognition systems or text
generation tasks where the exact wording is important. While it can indicate errors in the generated response,
it doesn't directly tell us whether the answer provided is factually correct or helpful in addressing the help
desk question. For help desk applications, a more semantic understanding of correctness is needed, which F1-
score can offer.

In conclusion, the F1 score is the most suitable metric because it comprehensively evaluates the accuracy of
the fine-tuned LLM by considering both precision and recall, which are essential for a successful help desk
application.

Relevant Link:

Precision and Recall

F1 Score

Question: 97
A company is using Retrieval Augmented Generation (RAG) with Amazon Bedrock and Stable Diffusion to generate
product images based on text descriptions. The results are often random and lack specific details. The company
wants to increase the specificity of the generated images.

Which solution meets these requirements?

A.Increase the number of generation steps.

B.Use the MASK_IMAGE_BLACK mask source option.
C.Increase the classifier-free guidance (CFG) scale.
D.Increase the prompt strength.

Answer: C

Explanation:

Here's a detailed justification for why option C, "Increase the classifier-free guidance (CFG) scale," is the most
appropriate solution to improve the specificity of generated product images using RAG with Amazon Bedrock
and Stable Diffusion:

The problem indicates that the generated images lack specificity and appear random. This suggests that the
text prompt is not strongly influencing the image generation process. Classifier-Free Guidance (CFG) scale
directly controls how much the image generation process is guided by the text prompt.

Increasing the CFG scale tells the Stable Diffusion model to pay more attention to the prompt when creating
the image. A higher CFG scale forces the model to adhere more closely to the textual description, thus
increasing the likelihood of specific details being reflected in the image and reducing randomness. It
 essentially amplifies the influence of the prompt over the diffusion process.

Option A, increasing generation steps, might slightly improve image quality but doesn't directly address the
specificity issue. More steps refine the generated image but doesn't change the image content's adherence to
the prompt's details.

Option B, using the MASK_IMAGE_BLACK mask source option, is relevant for in-painting scenarios (modifying
existing images) and not for generating images from scratch based on text descriptions. In-painting requires a
pre-existing image and is about changing specific regions, not guiding the overall generation.

Option D, increasing prompt strength, is a general concept. CFG scale is the specific parameter within Stable
Diffusion that effectively controls the prompt's influence during image generation. It's the direct way to
adjust "prompt strength".

Therefore, increasing the CFG scale will most effectively guide Stable Diffusion to produce images that are
more specific and closely aligned with the product descriptions used in the RAG system. This will reduce
randomness and improve the relevance of the generated product images.

Relevant resources:

Stable Diffusion Documentation: https://stability.ai/ (Look for documentation on CFG scale or guidance scale
within the Stable Diffusion model)
Hugging Face Diffusers Library: https://huggingface.co/docs/diffusers/using-
diffusers/conditional_image_generation (Provides examples and explanations of CFG scale usage in diffusion
models)

Question: 98
A company wants to implement a large language model (LLM) based chatbot to provide customer service agents
with real-time contextual responses to customers' inquiries. The company will use the company's policies as the
knowledge base.

Which solution will meet these requirements MOST cost-effectively?

A.Retrain the LLM on the company policy data.

B.Fine-tune the LLM on the company policy data.
C.Implement Retrieval Augmented Generation (RAG) for in-context responses.
D.Use pre-training and data augmentation on the company policy data.

Answer: C

Explanation:

The most cost-effective solution for providing customer service agents with real-time contextual responses
using a large language model (LLM) and company policies as a knowledge base is Retrieval Augmented
Generation (RAG).

Here's why:

RAG leverages existing LLMs: RAG doesn't require retraining or fine-tuning the LLM itself. This drastically
reduces the computational costs associated with training large models.

Real-time Contextualization: RAG dynamically retrieves relevant information from the company's policy data
in response to a customer's query. This allows the LLM to provide contextually accurate answers in real-time.

Knowledge Base Updates: Updating the knowledge base (company policies) in a RAG system is much simpler
 than retraining or fine-tuning an LLM. New policies can be added to the knowledge base, and the RAG system
will automatically incorporate them.

Cost Efficiency: Retraining or fine-tuning large LLMs is computationally expensive and time-consuming,
requiring significant infrastructure and expertise. RAG avoids these costs by using an existing LLM and
focusing on efficient information retrieval.

Alternatives are less suitable: Retraining the LLM (A) is overkill and costly for incorporating policy data. Fine-
tuning (B) is also more expensive than RAG. Pre-training (D) is irrelevant since an LLM is already available.
Data augmentation is not applicable here.

In summary, RAG provides a cost-effective, flexible, and scalable solution for integrating company policies
into a chatbot while leveraging the power of pre-trained LLMs.

Supporting Information:

Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks: https://arxiv.org/abs/2005.11401

AWS Documentation on RAG Implementation: Consult AWS documentation for services like Amazon Kendra,
Amazon OpenSearch Service and Amazon SageMaker for more details on implementing RAG solutions.

Question: 99
A company wants to create a new solution by using AWS Glue. The company has minimal programming experience
with AWS Glue.

Which AWS service can help the company use AWS Glue?

A.Amazon Q Developer
B.AWS Config
C.Amazon Personalize
D.Amazon Comprehend

Answer: A

Explanation:

The correct answer is A, Amazon Q Developer. Here's why:

Amazon Q Developer is designed to assist developers in various coding-related tasks, including those
involving AWS services like AWS Glue. Its capabilities extend to generating code snippets, providing
explanations of existing code, and debugging. This is particularly helpful for users with limited programming
experience because it can effectively lower the barrier to entry for using AWS Glue and other AWS services.
Amazon Q helps automate and simplify many of the tasks that would otherwise require a significant amount
of manual coding effort. For instance, it can assist in generating the necessary PySpark scripts for AWS Glue
jobs.

AWS Config focuses on assessing, auditing, and evaluating the configurations of AWS resources. While
valuable for governance and compliance, it doesn't directly aid in code generation or simplification for AWS
Glue.

Amazon Personalize is a machine learning service for real-time personalization and recommendation, a
separate domain. It does not assist directly with code generation or guidance on how to use AWS Glue.

Amazon Comprehend is a natural language processing (NLP) service for extracting insights from text data.
Again, this service is distinct and does not provide direct support or code assistance for AWS Glue.
 Therefore, Amazon Q Developer, designed for code generation and providing explanations, is the most
appropriate choice to support a company with minimal programming experience to use AWS Glue. It directly
addresses the need for coding assistance, making the process of utilizing AWS Glue more accessible.

Supporting links:

AWS Glue: https://aws.amazon.com/glue/

Amazon Q: https://aws.amazon.com/amazon-q/

Question: 100
A company is developing a mobile ML app that uses a phone's camera to diagnose and treat insect bites. The
company wants to train an image classification model by using a diverse dataset of insect bite photos from
different genders, ethnicities, and geographic locations around the world.

Which principle of responsible AI does the company demonstrate in this scenario?

A.Fairness
B.Explainability
C.Governance
D.Transparency

Answer: A

Explanation:

The company in this scenario is demonstrating the principle of Fairness in responsible AI development.
Fairness, in the context of AI, aims to ensure that AI systems do not create or reinforce unfair biases and that
their outcomes are equitable across different demographic groups. The company is actively collecting a
diverse dataset that encompasses photos from different genders, ethnicities, and geographic locations. This
conscious effort to gather diverse data is directly aimed at mitigating potential biases in the trained image
classification model.

If the model were trained only on images from a single demographic group (e.g., a specific ethnicity or
geographic location), it might perform poorly on images from other groups, leading to inaccurate diagnoses
for individuals outside the dominant demographic. This would result in an unfair and potentially harmful
outcome. By explicitly seeking data from different genders, ethnicities, and regions, the company is actively
working to prevent such biases and ensure that the model's diagnoses are accurate and reliable for all users,
regardless of their background. This aligns directly with the goal of fairness in AI systems. The other options
don't fit as well: Explainability refers to understanding how an AI model arrives at its decisions, Governance
relates to the organizational structures and processes for responsible AI development and deployment, and
Transparency focuses on making the AI system's behavior and limitations clear to users. While all these
principles are important, the scenario's emphasis on diverse data collection points directly towards
addressing the potential for unfair or biased outcomes, making fairness the most appropriate answer.

For further reading on Fairness in AI, you can refer to resources such as:

Google AI Principles: https://ai.google/principles/

Microsoft's Responsible AI Principles: https://www.microsoft.com/en-us/ai/responsible-ai
IBM's AI Ethics: https://www.ibm.com/watson/trustworthy-ai

Question: 101
A company is developing an ML model to make loan approvals. The company must implement a solution to detect
bias in the model. The company must also be able to explain the model's predictions.

Which solution will meet these requirements?

A.Amazon SageMaker Clarify

B.Amazon SageMaker Data Wrangler
C.Amazon SageMaker Model Cards
D.AWS AI Service Cards

Answer: A

Explanation:

The correct answer is A, Amazon SageMaker Clarify, because it directly addresses the requirements of
detecting bias and explaining model predictions. SageMaker Clarify helps identify potential bias in machine
learning models throughout the model development lifecycle, from data preparation to model deployment. It
achieves this by providing tools to analyze datasets for bias before training, during training, and after model
deployment. Clarify generates detailed reports on bias metrics for various subgroups within the dataset,
allowing developers to pinpoint specific areas of concern. Moreover, SageMaker Clarify offers explainability
capabilities that help understand which features contribute most significantly to a model's predictions. It uses
techniques like SHAP (SHapley Additive exPlanations) to assign importance scores to each feature for
individual predictions, enhancing model transparency.

Option B, Amazon SageMaker Data Wrangler, is primarily focused on data preparation and feature
engineering. While it is a crucial step in the ML pipeline, it doesn't inherently provide bias detection or model
explainability features. Option C, Amazon SageMaker Model Cards, is a documentation tool that allows users
to record information about their models, such as training data, performance metrics, and intended use. While
useful for model governance, it doesn't automatically detect bias or explain predictions. Option D, AWS AI
Service Cards, is a conceptual framework for documenting ethical considerations of AI services, but it doesn't
provide the technical capabilities to detect bias or explain predictions within a custom ML model.

Therefore, SageMaker Clarify is the most appropriate choice as it directly offers bias detection and model
explainability, which are the explicit requirements outlined in the scenario. It enables the company to not only
identify and mitigate bias in their loan approval model but also understand the reasons behind the model's
decisions, promoting fairness and transparency.

Supporting links:

Amazon SageMaker Clarify: https://aws.amazon.com/sagemaker/clarify/

Amazon SageMaker Data Wrangler: https://aws.amazon.com/sagemaker/data-wrangler/
Amazon SageMaker Model Cards: https://aws.amazon.com/blogs/machine-learning/use-amazon-sagemaker-
model-cards-to-enable-model-governance/

Question: 102
A company has developed a generative text summarization model by using Amazon Bedrock. The company will use
Amazon Bedrock automatic model evaluation capabilities.

Which metric should the company use to evaluate the accuracy of the model?

A.Area Under the ROC Curve (AUC) score

B.F1 score
C.BERTScore
D.Real world knowledge (RWK) score
 Answer: C

Explanation:

The correct answer is C. BERTScore. Here's why:

Generative text summarization models, as used with Amazon Bedrock, produce text outputs. Evaluating the
accuracy of such models requires metrics that assess the semantic similarity and quality of the generated
summaries compared to reference summaries or original documents.

BERTScore excels in this scenario because it leverages contextual embeddings from models like BERT to
compute a similarity score between the generated text and reference text. It focuses on semantic matching
rather than simple word overlap, capturing meaning even when different words are used to convey the same
information. This is crucial for summarization, where paraphrasing is expected.

Options A, B, and D are less suitable:

A. Area Under the ROC Curve (AUC) score: AUC is a measure of performance for binary classification
problems, not text generation. It assesses the model's ability to distinguish between two classes.
B. F1 score: F1 score, while used in information retrieval, focuses primarily on precision and recall based on
exact word matches. It doesn't effectively capture semantic similarity in generated text.
D. Real-world knowledge (RWK) score: RWK score evaluates the ability of a model to provide truthful
answers relative to external real-world knowledge. It does not help evaluate the accuracy of text
summarization.

Amazon Bedrock's automatic model evaluation often supports metrics relevant to the task. For generative
models, BERTScore provides a more nuanced and accurate reflection of summary quality than basic metrics
like ROUGE, which focuses on n-gram overlap. Using BERTScore helps the company ensure that the
generated summaries are both accurate and semantically similar to the source material.

Authoritative links for further research:

BERTScore Paper: https://arxiv.org/abs/1904.09675

Amazon Bedrock Documentation: (Search for automatic evaluation capabilities within the service)
https://aws.amazon.com/bedrock/

Question: 103
An AI practitioner wants to predict the classification of flowers based on petal length, petal width, sepal length,
and sepal width.

Which algorithm meets these requirements?

A.K-nearest neighbors (k-NN)

B.K-mean
C.Autoregressive Integrated Moving Average (ARIMA)
D.Linear regression

Answer: A

Explanation:

Here's a detailed justification for why K-Nearest Neighbors (k-NN) is the best choice for classifying flowers
based on petal and sepal measurements, and why the other options are unsuitable:

Justification for K-Nearest Neighbors (k-NN):

 k-NN is a supervised machine learning algorithm primarily used for classification and regression tasks. In this
scenario, we want to classify flowers, making classification the appropriate use case. k-NN is a non-
parametric, instance-based learning algorithm. It classifies a new data point (flower) based on the majority
class among its 'k' nearest neighbors in the feature space (petal length, petal width, sepal length, sepal
width).

Because we have labeled data (historical data associating specific flower measurements to specific flower
species), k-NN can "learn" from these examples. When presented with a new flower's measurements, the
algorithm identifies the 'k' most similar flowers in the training data based on distance metrics (e.g., Euclidean
distance). It then assigns the new flower to the class that is most prevalent among those 'k' neighbors.

k-NN is relatively simple to implement and understand, requiring minimal training. It is well-suited for multi-
class classification problems (like classifying multiple flower species). It is also robust to noisy data, although
performance can degrade with very high-dimensional datasets (which isn't the case here).

Why the other options are unsuitable:

K-Means: K-means is an unsupervised clustering algorithm. It aims to group data points into 'k' clusters based
on similarity. It doesn't classify based on prior labeled data. Since we need to classify (i.e., assign labels) the
flowers, unsupervised clustering is incorrect.

Autoregressive Integrated Moving Average (ARIMA): ARIMA is a time series forecasting model. It's designed
to predict future values based on past values in a sequence of data points indexed in time order. It's entirely
unrelated to flower classification based on physical characteristics.

Linear Regression: Linear regression is used to predict a continuous numerical value. While variations exist, it
isn't suitable for classification problems where the goal is to assign data points to distinct categories (flower
types). Linear regression would attempt to predict a continuous value related to flowers.

Authoritative Links:

AWS Machine Learning Documentation: https://aws.amazon.com/machine-learning/

Scikit-learn k-NN Documentation: https://scikit-
learn.org/stable/modules/generated/sklearn.neighbors.KNeighborsClassifier.html
Introduction to Machine Learning with Python (Andreas Müller & Sarah Guido): Excellent resource to
comprehend the core concepts.

In conclusion, given the specific goal of classifying flowers based on physical measurements with labeled
data, K-Nearest Neighbors is the most suitable algorithm among the choices. The other options address
different types of machine learning problems, making them inappropriate for this classification task.

Question: 104
A company is using custom models in Amazon Bedrock for a generative AI application. The company wants to use a
company managed encryption key to encrypt the model artifacts that the model customization jobs create.

Which AWS service meets these requirements?

A.AWS Key Management Service (AWS KMS)

B.Amazon Inspector
C.Amazon Macie
D.AWS Secrets Manager

Answer: A
 Explanation:

The correct answer is AWS Key Management Service (AWS KMS) because it directly addresses the
requirement of using a company-managed encryption key to encrypt model artifacts created by Amazon
Bedrock model customization jobs. AWS KMS allows users to create, manage, and control cryptographic keys
used to encrypt data at rest and in transit. It provides a centralized key management solution that integrates
with various AWS services, including Amazon Bedrock. Using KMS enables the company to maintain control
over the encryption keys, ensuring compliance with security and regulatory requirements.

Amazon Inspector focuses on identifying security vulnerabilities and deviations from best practices within an
AWS environment. Amazon Macie is a data security and data privacy service that uses machine learning and
pattern matching to discover and protect sensitive data in AWS. AWS Secrets Manager helps you manage,
retrieve, and rotate secrets, such as database credentials, API keys, and other sensitive information. While
these services provide valuable security functions, they do not directly address the requirement of providing
company-managed keys for encrypting Bedrock model artifacts. Only AWS KMS provides the functionality to
manage and control the encryption keys needed for this purpose. Specifically, you can configure Bedrock to
use a KMS key that you own and manage to encrypt the artifacts generated during the model customization
process. This provides an extra layer of security by ensuring that only those with access to the KMS key can
decrypt the model artifacts.

References:

AWS KMS
Encryption and security in Amazon Bedrock

Question: 105
A company wants to use large language models (LLMs) to produce code from natural language code comments.

Which LLM feature meets these requirements?

A.Text summarization
B.Text generation
C.Text completion
D.Text classification

Answer: B

Explanation:

Here's a detailed justification for why the correct answer is B (Text Generation) and why the other options are
less suitable, in the context of a company using LLMs to produce code from natural language comments:

The core requirement is to create new code based on input text (the code comments). Text Generation is
specifically designed for this purpose: to produce novel text, adhering to the patterns and knowledge it has
learned during its training. In this scenario, the LLM would ingest the comment and then generate
corresponding code as its output.

Let's examine why the other options are less effective:

A. Text Summarization: Text summarization aims to condense a longer piece of text into a shorter, more
concise version while retaining the key information. This is not what the company needs; they need to produce
code, not a shortened version of the comments.
C. Text Completion: While text completion can add to existing text, it often focuses on predicting the
immediate next word or phrase in a sequence. It might be helpful within a larger generation process, but it's
 not the primary function needed to translate full comments into full code blocks. Text completion is less
about creating entirely new content and more about extending existing content.
D. Text Classification: Text classification categorizes text into predefined classes or labels. It's useful for
sentiment analysis, topic detection, or spam filtering, but it doesn't generate any new text. It wouldn't be able
to produce any code from the natural language comment input.

In conclusion, Text Generation is the most appropriate LLM feature because its primary function is to create
new, meaningful text based on a given prompt, which perfectly aligns with the company's need to generate
code from natural language code comments.

Authoritative Links for further Research:

Amazon Bedrock Text Generation: https://aws.amazon.com/bedrock/ (Explore the capabilities of text

generation models on AWS)
Overview of Natural Language Processing: https://aws.amazon.com/machine-learning/nlp/ (A general
introduction to NLP concepts, including text generation)

Question: 106
A company is introducing a mobile app that helps users learn foreign languages. The app makes text more
coherent by calling a large language model (LLM). The company collected a diverse dataset of text and
supplemented the dataset with examples of more readable versions. The company wants the LLM output to
resemble the provided examples.

Which metric should the company use to assess whether the LLM meets these requirements?

A.Value of the loss function

B.Semantic robustness
C.Recall-Oriented Understudy for Gisting Evaluation (ROUGE) score
D.Latency of the text generation

Answer: C

Explanation:

The correct answer is C. Recall-Oriented Understudy for Gisting Evaluation (ROUGE) score.

Here's why:

ROUGE (Recall-Oriented Understudy for Gisting Evaluation): ROUGE is a set of metrics used for evaluating
automatic summarization and machine translation tasks. Because the company is improving text readability
by calling a large language model, ROUGE allows the LLM output to be measured against a human created
reference which will give the company an understanding on how similar and how useful the LLM output is to
the human created example.

Value of the loss function: The loss function measures the difference between the predicted output and the
actual output during training. While important for training, a low loss doesn't directly guarantee the output is
readable or resembles the examples. A model can overfit, achieving low loss on the training data but failing to
generalize to new, unseen text.

Semantic robustness: Semantic robustness assesses the model's ability to maintain the meaning of the
output even with small changes in the input. While important, it does not directly relate to the readability or
similarity to the company's provided examples.

Latency of the text generation: Latency measures the time taken to generate text. While important for user
experience, it doesn't assess the readability or quality of the generated text.
 In summary, ROUGE is the best option here because it is specifically designed to measure the similarity
between the LLM's output and the reference "more readable" versions provided by the company, focusing on
recall (how much of the reference text is captured by the LLM output) and thereby helping assess whether
the LLM meets the requirement of resembling the provided examples.

Supporting links:

ROUGE: https://en.wikipedia.org/wiki/ROUGE_(metric)
Evaluation metrics in NLP: https://towardsdatascience.com/evaluation-metrics-for-language-modeling-a-
guide-to-nlp-performance-evaluation-b331c02205f4

Question: 107
A company notices that its foundation model (FM) generates images that are unrelated to the prompts. The
company wants to modify the prompt techniques to decrease unrelated images.

Which solution meets these requirements?

A.Use zero-shot prompts.

B.Use negative prompts.
C.Use positive prompts.
D.Use ambiguous prompts.

Answer: B

Explanation:

The correct answer is B. Use negative prompts.

Here's why: Foundation Models (FMs), even sophisticated ones, can sometimes generate outputs that don't
align with the intended request. This is because the model might focus on unintended aspects of a prompt or
lack a precise understanding of the desired output. Negative prompting is a technique that directly addresses
this issue.

Negative prompts explicitly tell the FM what not to include in the generated output. By specifying elements or
characteristics to avoid, you guide the model away from generating unrelated images. In the given scenario, if
the FM is generating images with, say, artifacts or unintended objects, negative prompts can be used to
instruct the model to exclude these elements. This clarifies the desired output and reduces the chances of
irrelevant images being generated.

Let's examine why the other options are less effective:

A. Use zero-shot prompts: Zero-shot prompting involves prompting the model without providing any
examples. While useful, it doesn't inherently prevent unrelated image generation. It relies on the model's
existing knowledge, which might be insufficient.
C. Use positive prompts: Positive prompts describe what to include in the output. While essential, they might
not be enough to prevent the inclusion of unwanted elements. Positive prompts are more about guiding the
model towards the desired elements but don't provide exclusion rules.
D. Use ambiguous prompts: Ambiguous prompts, by definition, are unclear and will likely worsen the problem
of unrelated image generation. They provide the model with even more freedom to misinterpret the request,
leading to unpredictable results.

In summary, negative prompts are the most effective way to directly address the problem of FMs generating
unrelated images. By clearly specifying what should not be included, you refine the generation process and
increase the likelihood of obtaining the desired results.
 Supporting Concepts and Resources:

Prompt Engineering: The practice of designing effective prompts to elicit desired responses from large
language models or FMs. Negative prompting is a key technique in prompt engineering.
Foundation Models: Large AI models pre-trained on vast amounts of data that can be adapted for various
downstream tasks. They can be used to generate text, images, and other content.
Amazon Bedrock: A service that provides access to foundation models from various providers. This allows
users to choose a model that best suits their needs.

Further Reading:

While not a direct AWS documentation on AIF-C01, exploring prompt engineering techniques for large
language models provides context and insight into foundation models and their application.
https://www.promptingguide.ai/
Amazon Bedrock documentation on model customization: https://aws.amazon.com/bedrock/ (Look for
sections related to fine-tuning and prompt engineering). While this doesn't explicitly mention "negative
prompts" as a feature (they're implemented in the prompt itself, not as a bedrock feature), understanding
model customization helps understand the context.

Question: 108
A company wants to use a large language model (LLM) to generate concise, feature-specific descriptions for the
company’s products.

Which prompt engineering technique meets these requirements?

A.Create one prompt that covers all products. Edit the responses to make the responses more specific, concise,
and tailored to each product.
B.Create prompts for each product category that highlight the key features. Include the desired output format
and length for each prompt response.
C.Include a diverse range of product features in each prompt to generate creative and unique descriptions.
D.Provide detailed, product-specific prompts to ensure precise and customized descriptions.

Answer: B

Explanation:

The correct answer is B because it aligns with best practices for prompt engineering when dealing with
diverse product information and the need for both conciseness and feature specificity.

Here's a detailed justification:

Option A is inefficient and ineffective. Creating a single prompt for all products would likely result in generic
and less specific descriptions. Editing the responses post-generation is time-consuming and negates the
purpose of using an LLM for automation. It doesn't leverage the LLM's ability to tailor responses based on
specific inputs.

Option C is unsuitable because it prioritizes creativity and uniqueness, which are not the primary requirements
in this scenario. The company needs concise, feature-specific descriptions, not necessarily creative ones.
Introducing diverse features into each prompt could lead to unfocused and less relevant outputs.

Option D seems promising but is actually less scalable and manageable than Option B. While product-specific
prompts can yield precise descriptions, creating and maintaining prompts for every single product, especially
in a large product catalog, would be extremely resource-intensive and prone to errors.

Option B offers a balanced approach. By grouping products into categories and creating prompts that
 highlight key features within those categories, the company can leverage the LLM's ability to generalize
within a similar context. Specifying the desired output format and length in each prompt is crucial for
ensuring conciseness and consistency across all generated descriptions. This approach allows for tailored
descriptions that are still manageable from a prompt engineering perspective. It's more efficient and likely
more effective than creating individual prompts for each product. By leveraging the LLM's understanding of a
given category, the model can infer and highlight the important features, ensuring the model generates
descriptions that are both concise and informative.

Here are some authoritative resources on prompt engineering:

Prompt Engineering Guide: https://www.promptingguide.ai/

OpenAI's documentation on prompt engineering: https://platform.openai.com/docs/guides/prompt-
engineering

These resources provide a comprehensive overview of prompt engineering techniques, including strategies
for creating effective prompts, optimizing responses, and avoiding common pitfalls. They will help you further
understand why Option B is the most suitable choice in this scenario.

Question: 109
A company is developing an ML model to predict customer churn. The model performs well on the training dataset
but does not accurately predict churn for new data.

Which solution will resolve this issue?

A.Decrease the regularization parameter to increase model complexity.

B.Increase the regularization parameter to decrease model complexity.
C.Add more features to the input data.
D.Train the model for more epochs.

Answer: B

Explanation:

The scenario describes a classic case of overfitting. The model has learned the training data too well,
including its noise and specific patterns, leading to poor generalization on unseen data.

Option B, increasing the regularization parameter, addresses this issue directly. Regularization techniques
penalize overly complex models, preventing them from fitting the training data too closely. By increasing the
regularization strength, the model is forced to simplify, reducing its sensitivity to noise in the training data and
promoting better generalization.

Options A, C, and D might worsen the problem. Decreasing the regularization parameter (A) would increase
model complexity, further exacerbating overfitting. Adding more features (C), without careful feature
selection and engineering, can also increase complexity and the risk of overfitting. Training for more epochs
(D) might further refine the model's fit to the training data, potentially making it even more overfit to the initial
dataset.

Therefore, the most appropriate solution to improve the model's performance on new data is to increase the
regularization parameter, effectively decreasing the model's complexity and improving its ability to
generalize. Regularization reduces variance by increasing the bias which makes models more generalizable.

For more information on regularization techniques:

Regularization in Machine Learning: https://www.geeksforgeeks.org/regularization-in-machine-learning/

 Overfitting and Underfitting With Machine Learning Algorithms:
https://machinelearningmastery.com/overfitting-and-underfitting-with-machine-learning-algorithms/

Question: 110
A company is implementing intelligent agents to provide conversational search experiences for its customers. The
company needs a database service that will support storage and queries of embeddings from a generative AI
model as vectors in the database.

Which AWS service will meet these requirements?

A.Amazon Athena
B.Amazon Aurora PostgreSQL
C.Amazon Redshift
D.Amazon EMR

Answer: B

Explanation:

The correct answer is B. Amazon Aurora PostgreSQL.

Here's why:

Vector Storage and Search: Modern intelligent agents relying on generative AI often use embeddings (vector
representations of text, images, etc.) to perform semantic search and similarity comparisons. A database
capable of storing and efficiently querying these vector embeddings is crucial.

Amazon Aurora PostgreSQL's pgvector Extension: Aurora PostgreSQL supports the pgvector extension,
which provides native vector storage and similarity search capabilities. This allows storing embeddings
directly in the database and performing k-nearest neighbors (k-NN) searches or other vector-based queries
efficiently.

Suitability for Conversational AI: Aurora PostgreSQL is a relational database suitable for managing data for
customer interactions. The vector search capabilities enhance it for conversational experiences where
semantic similarity is key.

Other Options:

Amazon Athena: Athena is a serverless query service for data in Amazon S3. While it can query data stored in
S3, it's not designed for low-latency vector similarity searches required for conversational AI applications.

Amazon Redshift: Redshift is a data warehouse optimized for analytical queries on large datasets. While it
might be able to store vector data, it's not the best choice for real-time, low-latency semantic search. Also,
the pgvector support is limited compared to Aurora PostgreSQL.

Amazon EMR: EMR is a managed Hadoop service for big data processing. It's not a database service
optimized for vector storage and similarity search.

Why Aurora PostgreSQL is Ideal: Aurora PostgreSQL with pgvector is ideal because it offers a relational
database structure combined with specialized vector search capabilities. This enables efficient storage,
querying, and management of embeddings, supporting the real-time search requirements of conversational AI
applications.

Supporting Links:
 pgvector: https://github.com/pgvector/pgvector
Amazon Aurora PostgreSQL: https://aws.amazon.com/rds/aurora/postgresql/

Question: 111
A financial institution is building an AI solution to make loan approval decisions by using a foundation model (FM).
For security and audit purposes, the company needs the AI solution's decisions to be explainable.

Which factor relates to the explainability of the AI solution's decisions?

A.Model complexity
B.Training time
C.Number of hyperparameters
D.Deployment time

Answer: A

Explanation:

The correct answer is A, Model complexity. Explainability in AI, especially in high-stakes domains like loan
approval, refers to the ability to understand and justify the AI's decision-making process. Model complexity
directly impacts explainability.

Foundation Models (FMs), being extremely large and complex neural networks with billions or even trillions of
parameters, inherently present explainability challenges. The intricate relationships learned by these models
can be opaque, making it difficult to trace back from a decision to the specific input features and internal
calculations that led to it. A less complex model, such as a simpler decision tree or linear model, is easier to
understand because its decision boundaries and feature importance are more readily apparent.

Options B, C, and D are not directly related to explainability. Training time (B) refers to the duration required to
train the model. The number of hyperparameters (C) influences model performance, but not directly
explainability. Deployment time (D) pertains to the time taken to deploy the model. While optimizing these
factors is important for operational efficiency, they do not affect whether the model's decisions can be
understood.

Therefore, for a financial institution that requires explainable AI for loan approvals, managing model
complexity becomes crucial. Techniques like feature importance analysis, SHAP values, or using simpler
models (potentially sacrificing some accuracy) might be necessary to achieve the desired level of
transparency and auditability. More complex models often require specialized explainability methods, which
can add overhead and may not always provide complete insight.

Relevant Resource:

Explainable AI (XAI): https://www.ibm.com/topics/explainable-ai - Provides a general overview of XAI.

Question: 112
A pharmaceutical company wants to analyze user reviews of new medications and provide a concise overview for
each medication.

Which solution meets these requirements?

A.Create a time-series forecasting model to analyze the medication reviews by using Amazon Personalize.
B.Create medication review summaries by using Amazon Bedrock large language models (LLMs).
 C.Create a classification model that categorizes medications into different groups by using Amazon
SageMaker.
D.Create medication review summaries by using Amazon Rekognition.

Answer: B

Explanation:

The correct answer is B because Amazon Bedrock, with its Large Language Models (LLMs), is ideally suited
for summarizing text data like medication reviews. The requirement is to analyze user reviews and provide
concise overviews for each medication. LLMs are specifically designed for natural language processing (NLP)
tasks, including text summarization, sentiment analysis, and information extraction. Bedrock provides access
to various pre-trained LLMs from different providers, enabling the company to choose the most appropriate
model for their specific needs.

Option A (Amazon Personalize) is incorrect because it is primarily used for recommendation systems and
personalized experiences, not for analyzing and summarizing text. Personalize uses time-series forecasting to
predict user behavior, which is not relevant to this scenario.

Option C (Amazon SageMaker) is incorrect because while SageMaker can be used to create custom machine
learning models, it requires significant expertise in machine learning and would be overkill for a simple
summarization task. Using pre-trained LLMs in Bedrock offers a faster and more cost-effective solution.
Moreover, creating a classification model as described wouldn't provide concise summaries.

Option D (Amazon Rekognition) is incorrect because it is designed for image and video analysis, not text
processing. Rekognition can identify objects, people, and scenes in images and videos, which is irrelevant to
summarizing medication reviews.

Bedrock's LLMs can efficiently process the user reviews and generate concise, informative summaries
highlighting the key aspects of each medication, thereby directly addressing the company's requirements.
The speed of deployment and the reduced complexity compared to building a model in SageMaker are also
significant advantages.

Amazon Bedrock DocumentationAmazon SageMaker DocumentationAmazon Rekognition

DocumentationAmazon Personalize Documentation

Question: 113
A company wants to build a lead prioritization application for its employees to contact potential customers. The
application must give employees the ability to view and adjust the weights assigned to different variables in the
model based on domain knowledge and expertise.

Which ML model type meets these requirements?

A.Logistic regression model

B.Deep learning model built on principal components
C.K-nearest neighbors (k-NN) model
D.Neural network

Answer: A

Explanation:

Here's a detailed justification for why Logistic Regression is the best answer:
 Why Logistic Regression is Suitable:

Interpretability: Logistic regression models are inherently interpretable. The coefficients associated with
each feature (variable) directly represent the impact of that feature on the predicted probability of a lead
being high-potential.

Adjustable Weights: The coefficients in a logistic regression model act as weights for the variables. By
adjusting these coefficients, employees can directly influence how each variable contributes to the lead
prioritization score, incorporating their domain knowledge. This is easily accomplished using standard
statistical packages.

Direct Relationship: It models the probability of a binary outcome (e.g., high-potential vs. low-potential lead)
using a logistic function. This output is straightforward and intuitive to understand as a probability score.

Transparency: Unlike complex models such as neural networks, the decision-making process of logistic
regression is relatively transparent. It's easier to understand why a particular lead was assigned a certain
score. This transparency builds trust and allows for better calibration by the domain experts.

Suitability for Prioritization: Logistic Regression excels in ranking and prioritization tasks, allowing leads to
be ordered based on their predicted probability of conversion.

Why the Other Options are Less Suitable:

Deep learning model built on principal components: Deep learning models are often black boxes and lack
interpretability, even when PCA is employed. It's very difficult for users to adjust weights in a direct and
understandable manner. Principal Components Analysis (PCA) reduces dimensionality but can further obscure
the relationship between original features and the output.

K-nearest neighbors (k-NN) model: While k-NN can be simple to understand, it doesn't provide explicit
weights that users can easily adjust. The influence of each variable is determined by the distance metric and
the neighborhood structure, which are not as intuitive to manipulate.

Neural network: Neural networks are complex models that are notoriously difficult to interpret. Even with
techniques like feature importance, it's almost impossible to allow domain experts to directly adjust the
impact of individual variables in a meaningful way. They are also overkill for a simpler prioritization task.

Authoritative Links for Further Research:

Logistic Regression Explanation: https://towardsdatascience.com/understanding-logistic-regression-

9b02c0da1023

Model Interpretability: https://christophm.github.io/interpretable-ml-book/

Question: 114
HOTSPOT
-

A company wants to build an ML application.

Select and order the correct steps from the following list to develop a well-architected ML workload. Each step
should be selected one time.
Answer:
Explanation:

Step 1: Define business goal and frame ML problem.

Before starting with model development, it's essential to identify the business objective and define the ML
problem.

This involves determining what needs to be predicted or classified, collecting data requirements, and
understanding constraints.

Step 2: Develop model.

Once the problem is defined, the next step is to develop the model.

This includes data preprocessing, feature engineering, selecting ML algorithms, training models, and tuning
hyperparameters.

Step 3: Deploy model

 After training and optimizing the model, it needs to be deployed to a production environment.

This can involve containerization, setting up APIs, or integrating the model into an application for real-world
use.

Step 4: Monitor model.

After deployment, monitoring is crucial to ensure model performance does not degrade over time.

It involves tracking metrics such as accuracy, drift detection, and retraining the model when necessary.

Question: 115
Which strategy will determine if a foundation model (FM) effectively meets business objectives?

A.Evaluate the model's performance on benchmark datasets.

B.Analyze the model's architecture and hyperparameters.
C.Assess the model's alignment with specific use cases.
D.Measure the computational resources required for model deployment.

Answer: C

Explanation:

The most effective strategy for determining if a foundation model (FM) meets business objectives is to assess
the model's alignment with specific use cases (Option C). While benchmark datasets (Option A) provide a
general sense of model capability, they may not directly correlate to the actual business problem. A model
could perform well on a generic dataset but still fail to address the nuances of a particular business
application. Similarly, analyzing the model's architecture and hyperparameters (Option B) offers insights into
the model's design but doesn't guarantee it will effectively solve a business problem. Measuring
computational resources (Option D) is important for deployment and cost considerations, but it doesn't speak
to the model's efficacy in achieving business goals.

Alignment with specific use cases requires evaluating the model's performance on tasks that directly mirror
the intended business application. This involves testing the model with real-world data relevant to the
business, measuring key performance indicators (KPIs) that are meaningful to the business, and gathering
feedback from stakeholders who will be using the model's outputs. For example, if the FM is intended for
customer sentiment analysis, the assessment should involve testing its accuracy on customer reviews and
feedback specific to the business's products or services, and measuring metrics like precision, recall, and F1-
score for identifying positive, negative, and neutral sentiments. The business must then determine if the
model’s sentiment predictions are reliable enough to inform business decisions.

Furthermore, a business must evaluate the model's ability to integrate seamlessly into existing workflows and
systems. Consider factors like data compatibility, API accessibility, and integration with downstream
applications. A high-performing FM from a technical standpoint may be unusable if it doesn't align with the
 company's existing data infrastructure or workflow. Successful alignment means the model directly
contributes to achieving specific business objectives, such as increasing customer satisfaction, reducing
operational costs, or improving product development.

Useful resource for learning about evaluating model effectiveness:

AWS AI and Machine Learning Documentation: This is a great starting point to understand how AWS
approaches AI/ML solutions, including metrics for evaluation. https://aws.amazon.com/machine-learning/

Question: 116
A company needs to train an ML model to classify images of different types of animals. The company has a large
dataset of labeled images and will not label more data.

Which type of learning should the company use to train the model?

A.Supervised learning
B.Unsupervised learning
C.Reinforcement learning
D.Active learning

Answer: A

Explanation:

The correct answer is A: Supervised learning. Here's why:

Supervised learning is a machine learning paradigm where an algorithm learns from a labeled dataset. A
labeled dataset contains input features (in this case, images of animals) and corresponding output labels (the
type of animal). The model learns to map the inputs to the outputs. Because the company has a large, labeled
dataset of animal images with their corresponding animal types, supervised learning is the most appropriate
approach. The model will learn to associate the visual features in the images with the correct animal category
based on the existing labels.

Unsupervised learning, on the other hand, deals with unlabeled data. The algorithm tries to find patterns or
structures in the data without any prior knowledge of the desired output. This isn't suitable here as the goal is
to classify images into specific, predefined categories.

Reinforcement learning involves training an agent to make decisions in an environment to maximize a reward.
While potentially applicable in niche image recognition contexts, it's generally not the go-to solution when a
large labeled dataset is available. It involves an iterative process of trial and error, which is unnecessary given
the presence of labeled training data.

Active learning is a type of supervised learning where the algorithm selectively queries a human or oracle for
labels for the most informative data points. This aims to achieve high accuracy with fewer labeled examples.
The question specifies that the company will not label more data. Thus, active learning is not the correct
method here. The company already has a sufficiently large labeled dataset to perform supervised learning.

In summary, supervised learning allows the model to learn from the established relationships between image
features and animal labels in the existing dataset, which is exactly what the company requires to classify
images of different animal types effectively.

Further resources:

Supervised Learning: https://www.coursera.org/lecture/machine-learning/supervised-learning-GEY7c

AWS AI/ML: https://aws.amazon.com/machine-learning/
Question: 117
Which phase of the ML lifecycle determines compliance and regulatory requirements?

A.Feature engineering
B.Model training
C.Data collection
D.Business goal identification

Answer: D

Explanation:

The correct answer is D, Business goal identification. Here's why:

The business goal identification phase is crucial for defining the ML project's objectives, scope, and success
metrics. This initial phase sets the foundation for the entire ML lifecycle and directly influences subsequent
stages. Determining compliance and regulatory requirements is intrinsically linked to understanding the
business goals.

For example, if the goal is to build a model for financial fraud detection, compliance with regulations like KYC
(Know Your Customer) and anti-money laundering laws becomes paramount. Similarly, in the healthcare
sector, HIPAA compliance must be addressed if the model uses patient data. The business goals dictate the
type of data used, the features engineered, the model trained, and the overall deployment strategy.

Data collection (C), feature engineering (A), and model training (B) are all important phases, but they are
downstream processes that are informed by the upfront identification of business needs and associated
regulatory demands. The business goal sets the boundaries within which these later phases operate. If
compliance and regulatory requirements are not identified upfront, significant rework might be needed later
in the ML lifecycle, leading to project delays and increased costs. Only by first understanding the business
need can compliance with applicable laws and regulations be effectively determined. Ignoring this phase
jeopardizes legal viability and ethical use.

Refer to AWS documentation on machine learning lifecycle best practices for additional information:
https://docs.aws.amazon.com/whitepapers/latest/machine-learning-engineering-data-science/the-machine-
learning-development-process.html

Question: 118
A food service company wants to develop an ML model to help decrease daily food waste and increase sales
revenue. The company needs to continuously improve the model's accuracy.

Which solution meets these requirements?

A.Use Amazon SageMaker and iterate with newer data.

B.Use Amazon Personalize and iterate with historical data.
C.Use Amazon CloudWatch to analyze customer orders.
D.Use Amazon Rekognition to optimize the model.

Answer: A

Explanation:
 The correct answer is A, utilizing Amazon SageMaker and iterative model updates with new data. Here's why:

Amazon SageMaker's Comprehensive ML Lifecycle Support: SageMaker is a fully managed service that
covers the entire machine learning lifecycle, from data preparation and model building to training,
deployment, and monitoring. https://aws.amazon.com/sagemaker/
Model Building and Training: SageMaker provides a managed environment for building and training ML
models using various algorithms and frameworks. The food service company can use SageMaker to build a
model that predicts food demand based on various factors (day of the week, weather, etc.).
Iterative Improvement: The key requirement is continuous improvement. SageMaker facilitates retraining
models with newer data. As the company collects daily data about food waste and sales, this data can be fed
back into the SageMaker environment to refine and improve the accuracy of the model over time. This
iterative process is central to achieving the desired outcome.
Reducing Food Waste and Increasing Revenue: By accurately predicting food demand, the model can help
the company avoid over-preparation, thus reducing food waste. Simultaneously, it can help optimize inventory
and preparation to meet demand, potentially increasing sales revenue.
Amazon Personalize Inappropriateness: Amazon Personalize (option B) is primarily focused on delivering
personalized recommendations. While potentially useful for suggesting menu items, it is not the best choice
for predicting overall food demand and reducing waste. https://aws.amazon.com/personalize/
Amazon CloudWatch Inapplicability: Amazon CloudWatch (option C) is a monitoring service. While it can be
used to monitor the performance of the model, it doesn't directly contribute to model building or iterative
improvement. https://aws.amazon.com/cloudwatch/
Amazon Rekognition Unsuitability: Amazon Rekognition (option D) is an image and video analysis service. It's
not relevant to the problem of predicting food demand and reducing waste based on sales data.
https://aws.amazon.com/rekognition/
Continuous Learning: SageMaker supports techniques like online learning, where the model is continuously
updated with new data in real-time. This is ideal for adapting to changing customer preferences and external
factors.
Scalability and Cost-Effectiveness: SageMaker is scalable and cost-effective, allowing the company to
easily handle increasing data volumes and model complexity.
MLOps support: SageMaker has capabilities for managing the ML model in production.
In summary, SageMaker is the optimal choice as it is a comprehensive ML platform that facilitates model
building, training, deployment, and, most importantly, iterative improvement with newer data, addressing the
core requirement of continuously improving model accuracy for reducing food waste and increasing sales
revenue.

Question: 119
A company has developed an ML model to predict real estate sale prices. The company wants to deploy the model
to make predictions without managing servers or infrastructure.

Which solution meets these requirements?

A.Deploy the model on an Amazon EC2 instance.

B.Deploy the model on an Amazon Elastic Kubernetes Service (Amazon EKS) cluster.
C.Deploy the model by using Amazon CloudFront with an Amazon S3 integration.
D.Deploy the model by using an Amazon SageMaker endpoint.

Answer: D

Explanation:

The most suitable solution is deploying the model using an Amazon SageMaker endpoint. SageMaker provides
a fully managed service that simplifies deploying and scaling ML models. It eliminates the need for managing
 underlying infrastructure, such as servers, operating systems, and networking.

Option A (Amazon EC2) requires manual server management, including patching, scaling, and security
configurations, which contradicts the requirement of avoiding infrastructure management. Option B (Amazon
EKS) involves managing a Kubernetes cluster, which also demands considerable operational overhead. Option
C (Amazon CloudFront with Amazon S3) is primarily designed for content delivery and static website hosting,
not for deploying and serving ML models.

SageMaker endpoints automatically provision the necessary compute resources, handle scaling based on
prediction traffic, and provide monitoring and logging capabilities. The service integrates with other AWS
services, making it easy to build a complete ML pipeline. By using SageMaker, the company can focus on
model performance and business value instead of infrastructure concerns. It abstracts away the complexities
of server management, enabling rapid deployment and efficient scaling of the real estate sales price
prediction model.

Furthermore, SageMaker offers various endpoint deployment options, including real-time inference and batch
transform, allowing the company to choose the most suitable approach for their prediction needs. It supports
different ML frameworks, ensuring flexibility in model
deployment.https://aws.amazon.com/sagemaker/https://docs.aws.amazon.com/sagemaker/latest/dg/deploy-
model.html

Question: 120
A company wants to develop an AI application to help its employees check open customer claims, identify details
for a specific claim, and access documents for a claim.

Which solution meets these requirements?

A.Use Agents for Amazon Bedrock with Amazon Fraud Detector to build the application.
B.Use Agents for Amazon Bedrock with Amazon Bedrock knowledge bases to build the application.
C.Use Amazon Personalize with Amazon Bedrock knowledge bases to build the application.
D.Use Amazon SageMaker to build the application by training a new ML model.

Answer: B

Explanation:

The correct answer is B: Use Agents for Amazon Bedrock with Amazon Bedrock knowledge bases to build the
application.

Here's why: The problem requires an AI application that can perform several tasks: check customer claims,
identify details, and access documents related to a specific claim. This necessitates retrieval of information
from a structured knowledge base or document repository and then acting upon that information.

Agents for Amazon Bedrock: This is designed to orchestrate interactions with large language models (LLMs)
and connect them to external systems or knowledge sources. An agent can be configured to understand user
requests, query the necessary data sources, and take actions based on the results, making it ideal for
accessing claims information and initiating related tasks.

Amazon Bedrock Knowledge Bases: These provide a managed service for connecting LLMs to company
knowledge bases. This allows the agent to retrieve the relevant claim details and documents when processing
a claim. The LLM can then synthesize this information and present it to the employee.

Option A is incorrect because Amazon Fraud Detector is focused on identifying fraudulent activities, which is
not directly applicable to the stated requirements of accessing and managing customer claims. Option C uses
 Amazon Personalize, which is designed for recommendation systems and personalized experiences, not for
retrieving information and performing actions on existing data. Option D suggests building a new ML model
from scratch using Amazon SageMaker. While this is a powerful approach, it's overkill for this scenario, which
can be efficiently handled by leveraging pre-trained LLMs in Bedrock with a knowledge base. Furthermore,
building a new model would require substantial time and resources for data collection, training, and
deployment.

Using Agents for Amazon Bedrock in conjunction with Amazon Bedrock knowledge bases provides a suitable
solution to accomplish the tasks and requirements described in the prompt.

Supporting resources:

Agents for Amazon Bedrock

Amazon Bedrock Knowledge Base

Question: 121
A manufacturing company uses AI to inspect products and find any damages or defects.

Which type of AI application is the company using?

A.Recommendation system
B.Natural language processing (NLP)
C.Computer vision
D.Image processing

Answer: C

Explanation:

The correct answer is C, Computer Vision. Here's a detailed justification:

Computer vision is a field of artificial intelligence that enables computers to "see" and interpret images. In the
context of a manufacturing company inspecting products for damages and defects, computer vision is the
ideal AI application. It involves using algorithms to analyze images captured by cameras or sensors,
identifying patterns and anomalies indicative of defects. This automated inspection process can significantly
improve quality control, reduce manual labor, and increase efficiency compared to traditional human
inspection. The system can be trained on a dataset of images showing both acceptable and defective
products, enabling it to learn the characteristics of each category.

Option A, Recommendation systems, are designed to suggest items or content based on user preferences or
historical data. This is not applicable to identifying physical defects in manufactured goods. Option B, Natural
Language Processing (NLP), focuses on enabling computers to understand and process human language. This
is irrelevant to image-based defect detection. Option D, Image processing, while related, is more about
manipulating images (e.g., enhancing contrast, reducing noise) rather than extracting meaningful insights and
understanding what the image represents. Computer vision encompasses image processing but goes further
to provide a high-level understanding.

Therefore, the application described perfectly aligns with the capabilities of computer vision, which focuses
on extracting information and understanding from images to detect defects and damages in a manufacturing
setting. The other options are not relevant to the stated problem.

Further Research:

AWS Computer Vision Services: https://aws.amazon.com/machine-learning/computer-vision/

 Computer Vision Explained: https://www.ibm.com/topics/computer-vision

Question: 122
A company wants to create an ML model to predict customer satisfaction. The company needs fully automated
model tuning.

Which AWS service meets these requirements?

A.Amazon Personalize
B.Amazon SageMaker
C.Amazon Athena
D.Amazon Comprehend

Answer: B

Explanation:

Here's a detailed justification for why Amazon SageMaker is the correct answer, and why the others aren't,
considering the requirement for fully automated model tuning.

Amazon SageMaker is the most suitable choice because it provides comprehensive machine learning
capabilities, including automated model tuning through its AutoPilot feature. SageMaker AutoPilot
automatically explores different algorithm options, feature engineering techniques, and hyperparameter
configurations to find the best model for a given dataset. This fully automated process eliminates the need for
manual intervention, making it ideal for the company's requirements. AutoPilot streamlines the model
development process by automating tasks that typically require significant time and expertise from data
scientists.

Amazon Personalize is designed for building real-time personalized recommendations and doesn't focus on
general model building and tuning for customer satisfaction prediction. While Personalize offers some level of
automation, it's specific to recommendation systems and isn't a general-purpose ML platform for this
scenario.

Amazon Athena is a serverless interactive query service that enables analyzing data in Amazon S3 using SQL.
It's a great tool for data exploration and analysis but doesn't provide any machine learning model building or
tuning capabilities.

Amazon Comprehend is a natural language processing (NLP) service used for tasks like sentiment analysis
and entity recognition. While it can be used to process customer feedback, it doesn't offer automated model
tuning for creating a general predictive model for customer satisfaction based on multiple factors beyond just
text.

In summary, only Amazon SageMaker with its AutoPilot feature offers the complete, automated model tuning
solution required for building a predictive model for customer satisfaction.

Authoritative links:

Amazon SageMaker: https://aws.amazon.com/sagemaker/

Amazon SageMaker AutoPilot: https://aws.amazon.com/sagemaker/autopilot/

Question: 123
Which technique can a company use to lower bias and toxicity in generative AI applications during the post-
processing ML lifecycle?

A.Human-in-the-loop
B.Data augmentation
C.Feature engineering
D.Adversarial training

Answer: A

Explanation:

The correct answer is A: Human-in-the-loop (HITL). HITL is crucial in post-processing generative AI outputs to
mitigate bias and toxicity because it leverages human judgment to filter and refine generated content.
Generative AI models, while powerful, can sometimes produce biased or toxic content due to biases present in
their training data or inherent limitations in their algorithms.

Data augmentation (B) focuses on increasing the dataset size and variability during the training phase, not
specifically on addressing post-processing bias. Feature engineering (C) involves selecting and transforming
input features for better model performance; it is also primarily used during the model training phase.
Adversarial training (D) aims to make models more robust against adversarial attacks, a different concern
than inherent bias and toxicity.

HITL offers a direct and targeted intervention strategy in the post-processing stage. Human reviewers can
assess generated content, identify biased or offensive outputs, and make necessary corrections or flag
problematic instances for further model refinement. This allows for immediate feedback and adjustments,
thereby reducing the impact of biased or toxic content on end-users.

Using HITL in AWS involves services like Amazon Augmented AI (Amazon A2I), which enables integration of
human review workflows into machine learning applications. This service allows you to easily route
predictions with low confidence scores or predictions that trigger certain criteria (e.g., potentially toxic
content) to human reviewers. The reviewers can then validate or correct the predictions, providing valuable
feedback to improve the model's performance over time and mitigate harmful biases.

In conclusion, while data augmentation, feature engineering, and adversarial training play vital roles in AI
model development, only HITL directly addresses bias and toxicity within the generative AI application's
output in the post-processing phase by leveraging human intervention for content refinement and validation.

References:

Amazon Augmented AI (Amazon A2I): https://aws.amazon.com/augmented-ai/

AI Fairness 360: https://aif360.mybluemix.net/ (Although not AWS specific, a good resource for general AI
Fairness information)

Question: 124
A bank has fine-tuned a large language model (LLM) to expedite the loan approval process. During an external
audit of the model, the company discovered that the model was approving loans at a faster pace for a specific
demographic than for other demographics.

How should the bank fix this issue MOST cost-effectively?

A.Include more diverse training data. Fine-tune the model again by using the new data.
B.Use Retrieval Augmented Generation (RAG) with the fine-tuned model.
C.Use AWS Trusted Advisor checks to eliminate bias.
D.Pre-train a new LLM with more diverse training data.
 Answer: A

Explanation:

The most cost-effective solution to address the bias in the LLM's loan approval process is to augment the
training data with more diverse examples and re-fine-tune the model (Option A). This approach directly
tackles the root cause of the bias – a lack of representative data during the initial fine-tuning.

Here's why:

Directly Addresses Bias: The model's biased behavior likely stems from insufficient representation of certain
demographics in the training data. Adding more diverse data exposes the model to a wider range of scenarios
and patterns associated with different demographics, allowing it to learn more balanced decision-making.
Cost-Effective Fine-Tuning: Fine-tuning is computationally less expensive than pre-training a new LLM from
scratch (Option D). Fine-tuning adapts an existing, pre-trained model to a specific task, requiring fewer
resources and less time.
RAG is Not the Primary Solution: While Retrieval Augmented Generation (RAG) can provide the model with
additional context during inference, it primarily improves accuracy and relevance, not directly addresses the
inherent bias embedded during training. It's more of a patch than a cure in this scenario.
AWS Trusted Advisor is Irrelevant: AWS Trusted Advisor primarily focuses on cost optimization, security,
fault tolerance, and performance improvements for AWS resources. It does not provide bias detection or
mitigation tools for machine learning models. Therefore, Option C is not applicable.

By including more diverse training data and fine-tuning the model again, the bank can cost-effectively
mitigate the bias and ensure fairer loan approval decisions.

Supporting Links:

Fairness in Machine Learning: https://developers.google.com/machine-learning/fairness-overview

Amazon SageMaker Clarify: https://aws.amazon.com/sagemaker/clarify/ (While Clarify helps detect bias, the
primary fix is diverse data and retraining.)
Data Diversity in Machine Learning: Research on data diversity and its impact on model fairness.

Question: 125
HOTSPOT
-

A company has developed a large language model (LLM) and wants to make the LLM available to multiple internal
teams. The company needs to select the appropriate inference mode for each team.

Select the correct inference mode from the following list for each use case. Each inference mode should be
selected one or more times.
Answer:

Explanation:

1.Chatbot needing real-time predictions with minimal latency → Real-time inference .

A chatbot requires immediate responses to user queries.

Real-time inference is designed for low-latency predictions, making it the ideal choice.

This ensures that the chatbot can quickly understand and respond to user inputs without delays.

2.Processing gigabytes of text files on weekends → Batch transform.

A batch processing job does not require real-time responses.

Since this job processes large amounts of text on a schedule (weekends), batch inference is more efficient.

Batch transform allows processing large datasets at once, reducing costs and optimizing performance.

3.Engineering team creating an API for small text inputs with low-latency predictions → Real-time inference.

An API that processes small text inputs requires real-time predictions to maintain a smooth user experience.

Real-time inference ensures quick response times for users making API requests.
Question: 126
A company needs to log all requests made to its Amazon Bedrock API. The company must retain the logs securely
for 5 years at the lowest possible cost.

Which combination of AWS service and storage class meets these requirements? (Choose two.)

A.AWS CloudTrail
B.Amazon CloudWatch
C.AWS Audit Manager
D.Amazon S3 Intelligent-Tiering
E.Amazon S3 Standard

Answer: AD

Explanation:

The correct answer is A and D.

A. AWS CloudTrail: CloudTrail is an AWS service that enables governance, compliance, operational auditing,
and risk auditing of your AWS account. By enabling CloudTrail, the company can log all API calls made to
Amazon Bedrock. This provides a detailed audit trail of who made which requests, when they were made, and
from where. CloudTrail is necessary for capturing the request logs themselves.
[https://aws.amazon.com/cloudtrail/]

D. Amazon S3 Intelligent-Tiering: Amazon S3 Intelligent-Tiering is a storage class designed to optimize

storage costs by automatically moving data to the most cost-effective access tier based on access patterns,
without operational overhead. Since the logs need to be retained for 5 years, and the access frequency is
likely to decrease over time, Intelligent-Tiering helps to move the logs to less expensive tiers like infrequent
access or glacier, providing cost savings. S3 offers durability and meets the security requirements due to its
access control policies and encryption options. [https://aws.amazon.com/s3/storage-classes/intelligent-
tiering/]

Why other options are incorrect:

B. Amazon CloudWatch: CloudWatch is primarily used for monitoring and observability. While it can collect
logs, it's not designed for long-term, cost-optimized archival of audit logs like CloudTrail. CloudWatch logs
can become expensive for long-term storage.

C. AWS Audit Manager: Audit Manager automates audit preparation by collecting evidence from your AWS
environment. It leverages services like CloudTrail but is not a replacement for it. It's a tool for audit readiness,
not the primary log collection service.

E. Amazon S3 Standard: While S3 Standard provides high durability and availability, it's the most expensive
S3 storage class. It's suitable for frequently accessed data, which isn't the expected use case for long-term
audit logs, making Intelligent-Tiering a more cost-effective choice.

Therefore, CloudTrail captures the Amazon Bedrock API request logs, and S3 Intelligent-Tiering provides a
secure, cost-optimized storage solution for retaining those logs for the required 5-year period.

Question: 127
An ecommerce company wants to improve search engine recommendations by customizing the results for each
user of the company’s ecommerce platform.
Which AWS service meets these requirements?

A.Amazon Personalize
B.Amazon Kendra
C.Amazon Rekognition
D.Amazon Transcribe

Answer: A

Explanation:

Amazon Personalize is a fully managed machine learning service that uses your data to generate item
recommendations for your users. It can also generate user segments based on the users' affinity for certain
items or item metadata.

https://docs.aws.amazon.com/personalize/latest/dg/what-is-personalize.html

Question: 128
A hospital is developing an AI system to assist doctors in diagnosing diseases based on patient records and
medical images. To comply with regulations, the sensitive patient data must not leave the country the data is
located in.

Which data governance strategy will ensure compliance and protect patient privacy?

A.Data residency
B.Data quality
C.Data discoverability
D.Data enrichment

Answer: A

Explanation:

A. Data residency

Data residency refers to the physical or geographic location where data is stored and processed. It's
especially important in healthcare and other regulated industries where laws (like HIPAA, GDPR, or local
health data protection acts) often require that sensitive personal data remain within a specific country or
region.

The hospital must ensure that patient data does not leave the country, making data residency the key data
governance strategy.

Question: 129
A company needs to monitor the performance of its ML systems by using a highly scalable AWS service.

Which AWS service meets these requirements?

A.Amazon CloudWatch
B.AWS CloudTrail
C.AWS Trusted Advisor
 D.AWS Config

Answer: A

Explanation:

A. Amazon CloudWatch

Amazon CloudWatch is a highly scalable AWS service designed to monitor performance metrics, logs, and
events for AWS resources, including ML systems.

It can:

Collect and track metrics (e.g., latency, error rates, throughput).

Monitor logs from applications (including those running ML models).

Create alarms to alert when performance thresholds are breached.

Enable dashboards to visualize system performance over time.

This makes it ideal for monitoring the performance of ML systems in production.

Question: 130
An AI practitioner is developing a prompt for an Amazon Titan model. The model is hosted on Amazon Bedrock. The
AI practitioner is using the model to solve numerical reasoning challenges. The AI practitioner adds the following
phrase to the end of the prompt: “Ask the model to show its work by explaining its reasoning step by step.”

Which prompt engineering technique is the AI practitioner using?

A.Chain-of-thought prompting
B.Prompt injection
C.Few-shot prompting
D.Prompt templating

Answer: A

Explanation:

Chain-of-thought promptingChain-of-thought prompting improves the reasoning ability of large language

models by prompting them to generate a series of intermediate steps that lead to the final answer of a multi-
step problem.

Question: 131
Which AWS service makes foundation models (FMs) available to help users build and scale generative AI
applications?

A.Amazon Q Developer
B.Amazon Bedrock
C.Amazon Kendra
D.Amazon Comprehend
 Answer: B

Explanation:

Amazon Bedrock is a fully managed service that provides access to a variety of high-performing foundation
models from leading AI companies and Amazon itself. It offers a single API for accessing different foundation
models, allowing for easy experimentation and integration into applications.

Question: 132
A company is building a mobile app for users who have a visual impairment. The app must be able to hear what
users say and provide voice responses.

Which solution will meet these requirements?

A.Use a deep learning neural network to perform speech recognition.

B.Build ML models to search for patterns in numeric data.
C.Use generative AI summarization to generate human-like text.
D.Build custom models for image classification and recognition.

Answer: A

Explanation:

A. Use a deep learning neural network to perform speech recognition

To support visually impaired users, the app must:

Understand what users say (speech-to-text)

Respond with voice output (text-to-speech)

These capabilities are provided by technologies based on deep learning neural networks, specifically:

Automatic Speech Recognition (ASR) – converts spoken language to text

Text-to-Speech (TTS) – converts text responses back to speech

Both are commonly built using deep learning techniques, which are particularly effective for modeling the
complex patterns in human speech.

Question: 133
A company wants to enhance response quality for a large language model (LLM) for complex problem-solving
tasks. The tasks require detailed reasoning and a step-by-step explanation process.

Which prompt engineering technique meets these requirements?

A.Few-shot prompting
B.Zero-shot prompting
C.Directional stimulus prompting
D.Chain-of-thought prompting

Answer: D
 Explanation:

Chain-of-thought (CoT) prompting is one of the oldest “chain of” methods for improving LLM performance – in
particular in the context of queries or tasks that need complex, human-like reasoning to reach an answer.

Question: 134
A company wants to keep its foundation model (FM) relevant by using the most recent data. The company wants to
implement a model training strategy that includes regular updates to the FM.

Which solution meets these requirements?

A.Batch learning
B.Continuous pre-training
C.Static training
D.Latent training

Answer: B

Explanation:

B. Continuous pre-trainingTo keep a foundation model (FM) updated with the most recent data on a regular
basis, you need a training approach that continually integrates new information. Continuous pre-training fits
this requirement because it periodically (or even continuously) retrains or fine-tunes the model with the latest
data, ensuring relevance and improved performance.Here's why the other options are less suitable:A. Batch
learning: Trains in large, discrete batches and may introduce significant delays between training cycles,
potentially causing the model to become stale.C. Static training: Trains the model once and does not update it
with new data, leading to outdated predictions.D. Latent training: Not a standard industry term or recognized
strategy for regularly updating foundation models.

Question: 135
HOTSPOT
-

A company wants to develop ML applications to improve business operations and efficiency.

Select the correct ML paradigm from the following list for each use case. Each ML paradigm should be selected
one or more times.
Answer:

Explanation:

1. Binary Classification → Supervised Learning.

Explanation: Binary classification involves predicting one of two possible classes (e.g., spam vs. not spam,
fraud vs. not fraud).
 Reasoning: It requires labeled data where the model learns from input-output pairs, making it a supervised
learning problem.

2. Multi-Class Classification → Supervised Learning .

Explanation: Multi-class classification extends binary classification to three or more classes (e.g., classifying
emails into "spam," "promotions," and "primary").

Reasoning: Since labeled data is needed to train the model, it is a supervised learning problem.

3. K-Means Clustering → Unsupervised Learning .

Explanation: K-means is a clustering algorithm that groups data points into clusters based on similarities.

Reasoning: It does not require labeled data but instead finds patterns and structures in the data, making it an
unsupervised learning method.

4. Dimensionality Reduction → Unsupervised Learning .

Explanation: Dimensionality reduction (e.g., PCA, t-SNE) reduces the number of features in a dataset while
retaining important information.

Reasoning: Since it identifies structure and patterns in the data without needing labels, it is an unsupervised
learning technique.

Question: 136
Which option is a characteristic of AI governance frameworks for building trust and deploying human-centered AI
technologies?

A.Expanding initiatives across business units to create long-term business value

B.Ensuring alignment with business standards, revenue goals, and stakeholder expectations
C.Overcoming challenges to drive business transformation and growth
D.Developing policies and guidelines for data, transparency, responsible AI, and compliance

Answer: D

Explanation:

D. Developing policies and guidelines for data, transparency, responsible AI, and compliance

An AI governance framework provides a structured approach to ensure that AI systems are:

Trustworthy

Ethical

Transparent

Compliant with laws and regulations

 These frameworks typically include policies and guidelines related to:

Data usage and privacy

Model transparency and explainability

Fairness and bias mitigation

Responsible AI development and deployment

Regulatory and legal compliance

This aligns directly with building human-centered AI technologies and fostering trust in AI.

Question: 137
An ecommerce company is using a generative AI chatbot to respond to customer inquiries. The company wants to
measure the financial effect of the chatbot on the company’s operations.

Which metric should the company use?

A.Number of customer inquiries handled

B.Cost of training AI models
C.Cost for each customer conversation
D.Average handled time (AHT)

Answer: C

Explanation:

C. Cost for each customer conversationTo measure the financial effect of a generative AI chatbot on an
ecommerce company’s operations, you want a metric that reflects the cost impact of each interaction. “Cost
for each customer conversation” captures how much the company is spending per inquiry handled by the
chatbot. This lets you directly compare the chatbot’s operational expenses versus human-agent costs or other
support channels.

Question: 138
A company wants to find groups for its customers based on the customers’ demographics and buying patterns.

Which algorithm should the company use to meet this requirement?

A.K-nearest neighbors (k-NN)

B.K-means
C.Decision tree
D.Support vector machine

Answer: B

Explanation:

K-meansThe company should use K-means to group customers based on demographics and buying patterns.
K-means is an unsupervised clustering algorithm that effectively partitions data into natural groups, making it
 ideal for discovering customer segments without prior labeling.

Question: 139
A company’s large language model (LLM) is experiencing hallucinations.

How can the company decrease hallucinations?

A.Set up Agents for Amazon Bedrock to supervise the model training.

B.Use data pre-processing and remove any data that causes hallucinations.
C.Decrease the temperature inference parameter for the model.
D.Use a foundation model (FM) that is trained to not hallucinate.

Answer: C

Explanation:

C. Decrease the temperature inference parameter for the model.

Hallucinations in large language models (LLMs) refer to situations where the model generates plausible-
sounding but false or inaccurate information.

One effective way to reduce hallucinations is to:

Lower the temperature parameter during inference

Temperature controls the randomness of model outputs.

A lower temperature (e.g., 0.2–0.5) makes the model more deterministic and conservative, leading to more
fact-based responses.

A higher temperature introduces more creativity and variation — which can increase hallucinations.

Question: 140
A company is using a large language model (LLM) on Amazon Bedrock to build a chatbot. The chatbot processes
customer support requests. To resolve a request, the customer and the chatbot must interact a few times.

Which solution gives the LLM the ability to use content from previous customer messages?

A.Turn on model invocation logging to collect messages.

B.Add messages to the model prompt.
C.Use Amazon Personalize to save conversation history.
D.Use Provisioned Throughput for the LLM.

Answer: B

Explanation:

Add messages to the model promptLarge language models (LLMs) typically rely on the context that is
provided directly in the input prompt when generating a response. To give the model the ability to use content
from previous customer messages, you need to include those past messages in the prompt for each new
 inference call. This approach ensures that the model has the necessary context to respond accurately based
on prior interactions.

Question: 141
A company’s employees provide product descriptions and recommendations to customers when customers call the
customer service center. These recommendations are based on where the customers are located. The company
wants to use foundation models (FMs) to automate this process.

Which AWS service meets these requirements?

A.Amazon Macie
B.Amazon Transcribe
C.Amazon Bedrock
D.Amazon Textract

Answer: C

Explanation:

C. Amazon Bedrock

Gives access to foundation models (from Anthropic, AI21, Meta, Cohere, Stability AI, and Amazon Titan) via API
without managing infrastructure.

Can be used to generate personalized responses, product descriptions, and recommendations based on
location and other context.

Question: 142
A company wants to upload customer service email messages to Amazon S3 to develop a business analysis
application. The messages sometimes contain sensitive data. The company wants to receive an alert every time
sensitive information is found.

Which solution fully automates the sensitive information detection process with the LEAST development effort?

A.Configure Amazon Macie to detect sensitive information in the documents that are uploaded to Amazon S3.
B.Use Amazon SageMaker endpoints to deploy a large language model (LLM) to redact sensitive data.
C.Develop multiple regex patterns to detect sensitive data. Expose the regex patterns on an Amazon
SageMaker notebook.
D.Ask the customers to avoid sharing sensitive information in their email messages.

Answer: A

Explanation:

A. Configure Amazon Macie to detect sensitive information in the documents that are uploaded to Amazon S3.

Uses machine learning and pattern matching to detect:

PII (names, emails, SSNs)

Financial data
 Credential leaks

No custom development needed — easy to configure with a few clicks.

Best fit for automating sensitive data detection with minimal effort.

Question: 143
HOTSPOT
-

A company is training its employees on how to structure prompts for foundation models.

Select the correct prompt engineering technique from the following list for each prompt template. Each prompt
engineering technique should be selected one time.

Answer:

Explanation:

1. "Classify the following text as either sports, politics, or entertainment: [input text]." : Zero-shot Learning .

Zero-shot learning refers to a model making predictions without having seen labeled examples beforehand.

Reasoning: In this case, the model is asked to classify text without any prior examples. It relies solely on its
pre-trained knowledge to determine the correct category.

2. "A [image 1], [image 2], and [image 3] are examples of [target class]. Classify the following image as [target
 class]." :Few-shot Learning .

Few-shot learning involves providing the model with a few labeled examples before asking it to make a
prediction.

Reasoning: The prompt provides three examples of a given class before asking the model to classify a new
instance. This aligns with few-shot learning, where the model generalizes from a small set of examples.

3. "[Question.] [Instructions to follow.] Think step by step and walk me through your thinking." :Chain-of-
Thought Reasoning .

Chain-of-thought reasoning is a prompting technique that encourages the model to break down complex
problems into step-by-step explanations.

Reasoning: The phrase "Think step by step and walk me through your thinking" explicitly instructs the model
to generate a structured, logical explanation rather than providing a direct answer. This is a characteristic of
chain-of-thought reasoning.

Question: 144
Which option is a benefit of using Amazon SageMaker Model Cards to document AI models?

A.Providing a visually appealing summary of a mode’s capabilities.

B.Standardizing information about a model’s purpose, performance, and limitations.
C.Reducing the overall computational requirements of a model.
D.Physically storing models for archival purposes.

Answer: B

Explanation:

The correct answer is B: Standardizing information about a model's purpose, performance, and limitations.

Amazon SageMaker Model Cards are designed to provide a standardized and centralized way to document
machine learning models. Their core benefit lies in establishing a consistent structure for recording key
information about a model throughout its lifecycle. This includes detailing the model's intended use case, its
performance metrics across various datasets, known limitations or biases, and ethical considerations.

Standardization ensures that different teams and stakeholders can easily understand and compare different
models. This is particularly important in regulated industries or organizations that require transparency and
accountability in their AI systems. Model Cards promote responsible AI practices by forcing a structured
consideration of potential risks and biases.

Option A, providing a visually appealing summary, while a desirable attribute, is not the primary purpose of
Model Cards. Visualizations might be part of a Model Card, but the focus is on comprehensive documentation.
Option C, reducing computational requirements, is unrelated to Model Cards. Model Cards are concerned with
 documentation, not model optimization. Option D, physically storing models, is also incorrect. Model Cards
store metadata and documentation about models, not the models themselves. Model artifacts are usually
stored in Amazon S3.

The value of Model Cards lies in facilitating model governance, auditability, and responsible AI practices by
centralizing and standardizing critical information about AI models. By making it easier to understand and
compare models, Model Cards helps to improve the overall quality and trustworthiness of AI systems.

Further research on Amazon SageMaker Model Cards can be found on the AWS documentation site:
https://docs.aws.amazon.com/sagemaker/latest/dg/model-cards.html and the AWS Machine Learning Blog:
https://aws.amazon.com/blogs/machine-learning/document-your-ml-models-with-amazon-sagemaker-model-
cards/.

Question: 145
What does an F1 score measure in the context of foundation model (FM) performance?

A.Model precision and recall

B.Model speed in generating responses
C.Financial cost of operating the model
D.Energy efficiency of the model’s computations

Answer: A

Explanation:

The F1 score, in the realm of evaluating foundation models (FMs), is a crucial metric that harmonizes precision
and recall into a single, consolidated value. Precision gauges the accuracy of positive predictions made by the
model, answering the question: "Out of all the items the model predicted as positive, how many were actually
positive?". Recall, conversely, measures the model's ability to identify all relevant instances, asking: "Out of
all the actual positive items, how many did the model correctly identify?".

The F1 score represents the harmonic mean of precision and recall. This means that it effectively balances
these two often-competing objectives. A high F1 score indicates that the model has both high precision and
high recall. In the context of FMs, this is particularly important. For instance, in a question-answering FM, high
precision means that when the model provides an answer, it's likely to be correct. High recall signifies that the
model is likely to find relevant answers when they exist.

Alternatives are incorrect because: Model speed, financial cost, and energy efficiency are important
considerations when deploying and managing FMs in a production environment but these are not directly
captured by the F1 score. The F1 score specifically focuses on the accuracy of the model's predictions in terms
of precision and recall. Options B, C, and D refer to cost, performance and operational efficiency, none of
which directly relate to the prediction accuracy aspect the F1 score is designed to measure.

In summary, the F1 score provides a holistic view of a foundation model's performance by capturing both its
accuracy (precision) and its completeness (recall), making it a valuable metric for model selection and
optimization.

Helpful Resources:

Wikipedia - F1 Score: https://en.wikipedia.org/wiki/F-score

Precision and Recall: https://developers.google.com/machine-learning/crash-course/classification/precision-
and-recall
Question: 146
A company deployed an AI/ML solution to help customer service agents respond to frequently asked questions.
The questions can change over time. The company wants to give customer service agents the ability to ask
questions and receive automatically generated answers to common customer questions.

Which strategy will meet these requirements MOST cost-effectively?

A.Fine-tune the model regularly.

B.Train the model by using context data.
C.Pre-train and benchmark the model by using context data.
D.Use Retrieval Augmented Generation (RAG) with prompt engineering techniques.

Answer: D

Explanation:

Here's a detailed justification for why option D (Use Retrieval Augmented Generation (RAG) with prompt
engineering techniques) is the most cost-effective strategy for the given scenario:

The problem highlights a dynamic FAQ system where questions change frequently. The goal is to enable
customer service agents to query the system and receive relevant, automatically generated answers. In this
context, Retrieval Augmented Generation (RAG) offers a superior approach to meet this requirement cost-
effectively.

RAG involves two primary stages: retrieval and generation. The retrieval stage searches a knowledge base (in
this case, the FAQ database) to find the most relevant information in response to the agent's query. The
generation stage then uses this retrieved information, combined with a pre-trained large language model
(LLM), to create a concise and informative answer.

Option A, fine-tuning the model regularly, is expensive. Fine-tuning requires significant computational
resources and a constantly updated, labeled dataset, which adds operational overhead. Moreover, retraining
the entire model for every change in FAQs is inefficient. Option B, training the model using context data,
suffers from similar drawbacks as fine-tuning, requiring large datasets and compute resources. Pre-training in
Option C is generally applicable to developing general-purpose models and is not designed for a company
with a specialized task.

RAG is much more cost-effective because it leverages a pre-trained LLM (which has already learned vast
amounts of general knowledge) and combines it with a specific knowledge base (the FAQs). New FAQs can be
added to the knowledge base without requiring extensive retraining of the LLM. Prompt engineering, a key
component of RAG, helps guide the LLM to generate the desired answer format and style.

Furthermore, RAG allows for efficient updates to the FAQ database. New questions and answers can be
indexed, making them available for retrieval almost immediately. This agility is crucial in a scenario where the
questions evolve over time. The computational cost of searching the knowledge base is generally much lower
than retraining a large language model.

In conclusion, RAG combines the power of pre-trained models with a dynamic, searchable knowledge base,
offering a cost-effective and adaptable solution for answering frequently asked questions. The ability to
update the knowledge base without extensive retraining and the use of prompt engineering for better output
makes RAG the most appropriate strategy.

For more information, you can explore these resources:

Retrieval Augmented Generation (RAG): https://aws.amazon.com/what-is/retrieval-augmented-generation/

Prompt Engineering: https://www.promptingguide.ai/
Question: 147
A company built an AI-powered resume screening system. The company used a large dataset to train the model.
The dataset contained resumes that were not representative of all demographics.

Which core dimension of responsible AI does this scenario present?

A.Fairness
B.Explainability
C.Privacy and security
D.Transparency

Answer: A

Explanation:

The scenario highlights a significant issue in AI development relating to biased datasets. The problem stems
from using a non-representative dataset for training the resume screening model.

Fairness is the most pertinent responsible AI dimension affected. A biased training dataset, as described,
inevitably leads to discriminatory outcomes. The AI model, learning from the skewed data, will likely
perpetuate and amplify existing biases present in the dataset, unfairly disadvantaging certain demographic
groups during the resume screening process. This contradicts the core principle of fairness, which dictates
that AI systems should treat all individuals and groups equitably, regardless of protected characteristics.

Explainability, while important, focuses on understanding how the model reaches its decisions. Privacy and
Security relate to data protection and system security, respectively, which are not explicitly the focus of the
scenario. Transparency refers to how well-documented and accessible the AI system and its data are. While a
biased dataset can impact transparency if the data source and composition aren't readily available, the root
cause of the problem as described is fundamentally about unfair outcomes derived from a skewed dataset.

Therefore, fairness is the core dimension of responsible AI violated in this scenario because the AI system is
likely to produce discriminatory results due to the biased training data. Addressing this would require
techniques to identify and mitigate bias in datasets and algorithms, ensuring equitable outcomes for all
candidates.

For more information on responsible AI and fairness, you can consult the following resources:

Google AI - Fairness: https://ai.google/responsibility/responsible-ai-practices/

Microsoft AI - Responsible AI: https://www.microsoft.com/en-us/ai/responsible-ai
IBM - AI Ethics: https://www.ibm.com/watson/ai-ethics

Question: 148
A global financial company has developed an ML application to analyze stock market data and provide stock
market trends. The company wants to continuously monitor the application development phases and to ensure that
company policies and industry regulations are followed.

Which AWS services will help the company assess compliance requirements? (Choose two.)

A.AWS Audit Manager

B.AWS Config
C.Amazon Inspector
D.Amazon CloudWatch
 E.AWS CloudTrail

Answer: AB

Explanation:

The correct answer is AB. Here's why:

A. AWS Audit Manager: AWS Audit Manager helps automate compliance assessments. It collects evidence
from AWS resources, evaluates it against predefined frameworks (such as industry regulations or company
policies), and provides audit-ready reports. This directly addresses the company's need to ensure compliance
with policies and regulations throughout the application development phases. https://aws.amazon.com/audit-
manager/

B. AWS Config: AWS Config continuously monitors and records AWS resource configurations and allows you
to automate the evaluation of recorded configurations against desired configurations. It provides a detailed
inventory of your AWS resources and tracks configuration changes over time, enabling the company to
continuously assess compliance. Config rules can be defined to check if resources are compliant with specific
configurations stipulated by industry regulations or internal company policies.
https://aws.amazon.com/config/

Why the other options are less suitable:

C. Amazon Inspector: Amazon Inspector is a vulnerability management service that automates security
vulnerability assessments. While important for security, it focuses on identifying software vulnerabilities in
EC2 instances and container images, not broader compliance with company policies and industry regulations
related to data handling and financial regulations as required by the company.

D. Amazon CloudWatch: Amazon CloudWatch is a monitoring and observability service that provides data and
actionable insights to monitor applications, respond to system-wide performance changes, optimize resource
utilization, and get a unified view of operational health. While it can be used to monitor some aspects related
to compliance, it doesn't directly provide tools for assessing compliance requirements in the way Audit
Manager and Config do.

E. AWS CloudTrail: AWS CloudTrail records API calls made on your AWS account and delivers log files to an
Amazon S3 bucket. This is crucial for auditing and security analysis. CloudTrail is extremely useful for
forensic analysis after a policy violation occurs but is more of a logging tool than an active compliance
assessment tool. While useful in an audit, CloudTrail on its own cannot determine if configurations and
resource usage are compliant.

In summary, AWS Audit Manager helps in assessing compliance against predefined frameworks and
generating audit reports, while AWS Config continuously monitors and evaluates resource configurations
against desired states. Both services are essential for a financial company to continuously monitor its ML
application development for adherence to company policies and industry regulations.

Question: 149
A company wants to improve the accuracy of the responses from a generative AI application. The application uses
a foundation model (FM) on Amazon Bedrock.

Which solution meets these requirements MOST cost-effectively?

A.Fine-tune the FM.

B.Retrain the FM.
C.Train a new FM.
 D.Use prompt engineering.

Answer: D

Explanation:

The correct answer is D. Use prompt engineering. Here's a detailed justification:

Prompt engineering is the practice of carefully crafting input prompts to guide a large language model (LLM),
like those used in Amazon Bedrock, towards generating desired outputs. It involves experimenting with
different wording, providing context, specifying the desired format, and even adding examples (few-shot
learning) within the prompt itself.

The key advantage of prompt engineering is its cost-effectiveness. It doesn't require retraining or fine-tuning
the underlying model. Fine-tuning (option A) and retraining (option B) are significantly more expensive and
time-consuming, involving adjusting the model's weights with new data. Training a completely new FM (option
C) is the most resource-intensive option.

For simply improving the accuracy of responses, particularly for a specific application, prompt engineering is
often the first and most efficient approach to try. By carefully designing the input, you can steer the model to
focus on relevant information and produce more accurate and relevant outputs without incurring the costs
associated with modifying the model itself. For example, you might add "Answer concisely in three sentences"
or "Focus on the following data points..." to the prompt. This focuses the FM and reduces extraneous or
inaccurate output. The prompt may need to be tweaked multiple times to reach optimal results.

The prompt serves as instructions for the foundation model, and a well-crafted prompt can significantly
impact the model's performance, allowing you to harness the power of generative AI in a cost-conscious
manner. Think of it like providing detailed instructions to a knowledgeable but easily distracted assistant.

Further resources for understanding prompt engineering and its application in cloud environments can be
found on the AWS documentation and AI/ML blogs. Some general information is available at:

AWS AI/ML Blog: https://aws.amazon.com/blogs/machine-learning/ (Search for prompt engineering articles)

Prompt Engineering Guide: https://www.promptingguide.ai/

Question: 150
A company wants to identify harmful language in the comments section of social media posts by using an ML
model. The company will not use labeled data to train the model.

Which strategy should the company use to identify harmful language?

A.Use Amazon Rekognition moderation.

B.Use Amazon Comprehend toxicity detection.
C.Use Amazon SageMaker built-in algorithms to train the model.
D.Use Amazon Polly to monitor comments.

Answer: B

Explanation:

The correct answer is B. Use Amazon Comprehend toxicity detection.

Here's why:

Amazon Comprehend provides pre-trained models that can analyze text for various insights, including toxicity
 detection. Toxicity detection specifically identifies the presence of inappropriate or harmful language in text.
Because the company will not be using labeled data to train a model, a pre-trained model is the best option.
Comprehend offers this functionality out-of-the-box, making it ideal for this scenario.

Option A, Amazon Rekognition moderation, focuses on image and video analysis, not text. Therefore, it's not
suitable for identifying harmful language in social media comments.

Option C, Amazon SageMaker built-in algorithms, could be used, but the problem description states that the
company will not be using labeled data. SageMaker built-in algorithms require labeled data. Therefore, this
would not be a fit.

Option D, Amazon Polly, is a text-to-speech service and doesn't analyze the content of the text for harmful
language. Its primary function is to convert text into spoken words.

Comprehend allows you to quickly analyze text without needing to train custom models, saving time and
resources. Its toxicity detection feature directly addresses the company's requirement to identify harmful
language in social media comments using a pre-built solution. The pre-trained nature of Comprehend's model
eliminates the need for labeled data, aligning perfectly with the problem description.

For further research, consult the official AWS documentation on Amazon Comprehend:
https://aws.amazon.com/comprehend/ and specifically the documentation for toxicity detection in
Comprehend.

Question: 151
A media company wants to analyze viewer behavior and demographics to recommend personalized content. The
company wants to deploy a customized ML model in its production environment. The company also wants to
observe if the model quality drifts over time.

Which AWS service or feature meets these requirements?

A.Amazon Rekognition
B.Amazon SageMaker Clarify
C.Amazon Comprehend
D.Amazon SageMaker Model Monitor

Answer: D

Explanation:

Here's a detailed justification for why Amazon SageMaker Model Monitor (Option D) is the most suitable
service for the media company's requirements:

The core need is to deploy a custom ML model and monitor its quality over time (detect model drift). Amazon
SageMaker is a comprehensive machine learning platform that provides tools for building, training, deploying,
and monitoring ML models. Specifically, SageMaker Model Monitor is designed to automatically detect and
remediate model drift in production.

SageMaker Model Monitor continuously monitors the quality of deployed machine learning models by
analyzing the input data and predictions. It detects deviations in model performance by comparing the current
model behavior to a baseline established during training. This helps identify data drift (changes in input data)
and concept drift (changes in the relationship between input and output).

Amazon Rekognition (Option A) is primarily for image and video analysis, not general model deployment and
monitoring. Amazon SageMaker Clarify (Option B) is focused on detecting bias in ML models, both before and
 after deployment, but it's not a continuous monitoring solution for model drift in the same way as Model
Monitor. Amazon Comprehend (Option C) is a natural language processing service, which is irrelevant to the
requirement of monitoring a deployed custom ML model for personalized content recommendations.

By using SageMaker Model Monitor, the media company can ensure the personalized content
recommendations remain accurate and relevant over time, even as viewer behavior and demographics evolve.
It helps proactively identify and address performance degradation, maintaining the effectiveness of the ML
model in production.

Further research:

Amazon SageMaker Model Monitor: https://aws.amazon.com/sagemaker/model-monitor/

Detect Model Drift: https://docs.aws.amazon.com/sagemaker/model-monitor/model-monitor-detect-model-
drift.html

Question: 152
A company is deploying AI/ML models by using AWS services. The company wants to offer transparency into the
models’ decision-making processes and provide explanations for the model outputs.

Which AWS service or feature meets these requirements?

A.Amazon SageMaker Model Cards

B.Amazon Rekognition
C.Amazon Comprehend
D.Amazon Lex

Answer: A

Explanation:

The correct answer is A, Amazon SageMaker Model Cards. Here's why:

The core requirement is transparency and explainability in AI/ML model decision-making. Amazon SageMaker
Model Cards are specifically designed to document model information, including its intended use,
performance metrics, training data, and ethical considerations. Crucially, they facilitate the creation of reports
detailing how a model arrives at its predictions, making it easier to understand and trust the model's output.
This addresses the need to provide explanations for model outputs.

Amazon Rekognition (B) is an image and video analysis service. While it can provide labels and bounding
boxes, it doesn't inherently offer detailed explanations for why it made those specific identifications in a
format useful for transparency regarding the model.

Amazon Comprehend (C) is a natural language processing (NLP) service used for tasks like sentiment analysis
and entity recognition. It provides insights into text data but doesn't directly provide transparency or
explainability on how the underlying model makes its decisions. Comprehend outputs results of its analysis,
not explanations of the model's inner workings.

Amazon Lex (D) is a service for building conversational interfaces. While it leverages ML, the focus is on
dialogue management and speech recognition, not on explaining the decision-making process of the
underlying models.

Therefore, only SageMaker Model Cards directly address the requirement of documenting and explaining
model behavior, making it the correct solution. It promotes accountability and trust by allowing stakeholders
to review the model's characteristics and understand its limitations.
 For further research, consult the official AWS documentation:

Amazon SageMaker Model Cards: https://aws.amazon.com/sagemaker/model-cards/

Question: 153
A manufacturing company wants to create product descriptions in multiple languages.

Which AWS service will automate this task?

A.Amazon Translate
B.Amazon Transcribe
C.Amazon Kendra
D.Amazon Polly

Answer: A

Explanation:

The correct answer is Amazon Translate. Amazon Translate is a neural machine translation service that
delivers fast, high-quality, and affordable language translation. It enables organizations to translate text and
documents across a wide range of languages, making it the ideal solution for the manufacturing company's
requirement to create product descriptions in multiple languages.

Here's why the other options are incorrect:

Amazon Transcribe: This service converts speech to text. It's useful for transcribing audio or video content,
not for translating text from one language to another.
Amazon Kendra: This is an intelligent search service powered by machine learning. It's used to index and
search through internal documents and data, not for language translation.
Amazon Polly: This service turns text into lifelike speech. It's the opposite of what the company needs; they
need text translated, not spoken.

Amazon Translate automates the translation process, reducing the manual effort and time required to create
product descriptions in different languages. It uses sophisticated machine learning models to ensure
accuracy and fluency in the translated text. The service integrates easily with other AWS services and can be
accessed via an API, making it simple to incorporate into existing workflows. The cost-effectiveness of
Amazon Translate is also a significant advantage, especially for businesses dealing with a large volume of
translation. The neural machine translation technology adapts over time, improving accuracy as more data is
processed. Therefore, for translating product descriptions, Amazon Translate is the most suitable AWS
service.

Further research:

Amazon Translate: https://aws.amazon.com/translate/

Question: 154
HOTSPOT
-

A company wants more customized responses to its generative AI models’ prompts.

Select the correct customization methodology from the following list for each use case. Each use case should be
selected one time.
Answer:

Explanation:

1. The models must be taught a new domain-specific task

Model fine-tuning.

When a model needs to adapt to a new task, especially within a specific domain (like legal, medical, or
finance), you typically:

Start with a pre-trained model.

Use labeled data from the new domain to fine-tune it.

This adapts the model’s behavior to the specific objectives or outputs of the domain-specific task.

Example: Fine-tuning a language model trained on general English to classify legal documents.

2. A limited amount of labeled data is available and more data is needed.

Data augmentation.

Data augmentation is used to:

 Artificially expand your dataset by generating modified versions of existing samples.

Improve model performance and generalization without requiring new labeled data.

Common in NLP, CV, and other domains — for instance:

Synonym replacement in NLP.

Image rotation or flipping in computer vision.

It’s especially useful when labeled data is scarce but you want to improve model robustness.

3. Only unlabeled data is available.

Continued pre-training.

When you have unlabeled data, you cannot fine-tune or augment it directly. Instead, you use the data to:

Continue pre-training a foundational model using self-supervised learning techniques (like masked language
modeling in NLP).

This helps the model adapt better to the target domain, even without labels.

Example: Using a corpus of technical manuals to further pre-train a language model before any fine-tuning.

Question: 155
Which AWS feature records details about ML instance data for governance and reporting?

A.Amazon SageMaker Model Cards

B.Amazon SageMaker Debugger
C.Amazon SageMaker Model Monitor
D.Amazon SageMaker JumpStart

Answer: U

Question: 156
A financial company is using ML to help with some of the company’s tasks.

Which option is a use of generative AI models?

A.Summarizing customer complaints

B.Classifying customers based on product usage
C.Segmenting customers based on type of investments
D.Forecasting revenue for certain products

Answer: U

Question: 157
A medical company wants to develop an AI application that can access structured patient records, extract relevant
information, and generate concise summaries.

Which solution will meet these requirements?

A.Use Amazon Comprehend Medical to extract relevant medical entities and relationships. Apply rule-based
logic to structure and format summaries.
B.Use Amazon Personalize to analyze patient engagement patterns. Integrate the output with a general
purpose text summarization tool.
C.Use Amazon Textract to convert scanned documents into digital text. Design a keyword extraction system to
generate summaries.
D.Implement Amazon Kendra to provide a searchable index for medical records. Use a template-based system
to format summaries.

Answer: U

Question: 158
A company is building an AI application to summarize books of varying lengths. During testing, the application fails
to summarize some books.

Why does the application fail to summarize some books?

A.The temperature is set too high.

B.The selected model does not support fine-tuning.
C.The Top P value is too high.
D.The input tokens exceed the model’s context size.

Answer: D

Question: 159
An airline company wants to build a conversational AI assistant to answer customer questions about flight
schedules, booking, and payments. The company wants to use large language models (LLMs) and a knowledge
base to create a text-based chatbot interface.

Which solution will meet these requirements with the LEAST development effort?

A.Train models on Amazon SageMaker Autopilot.

B.Develop a Retrieval Augmented Generation (RAG) agent by using Amazon Bedrock.
C.Create a Python application by using Amazon Q Developer.
D.Fine-tune models on Amazon SageMaker Jumpstart.

Answer: U

Question: 160
What is tokenization used for in natural language processing (NLP)?

A.To encrypt text data

B.To compress text files
 C.To break text into smaller units for processing
D.To translate text between languages

Answer: C

Question: 161
Which option is a characteristic of transformer-based language models?

A.Transformer-based language models use convolutional layers to apply filters across an input to capture local
patterns through filtered views.
B.Transformer-based language models can process only text data.
C.Transformer-based language models use self-attention mechanisms to capture contextual relationships.
D.Transformer-based language models process data sequences one element at a time in cyclic iterations.

Answer: C

Question: 162
A financial company is using AI systems to obtain customer credit scores as part of the loan application process.
The company wants to expand to a new market in a different geographic area. The company must ensure that it
can operate in that geographic area.

Which compliance laws should the company review?

A.Local health data protection laws

B.Local payment card data protection laws
C.Local education privacy laws
D.Local algorithm accountability laws

Answer: D

Question: 163
A company uses Amazon Bedrock for its generative AI application. The company wants to use Amazon Bedrock
Guardrails to detect and filter harmful user inputs and model-generated outputs.

Which content categories can the guardrails filter? (Choose two.)

A.Hate
B.Politics
C.Violence
D.Gambling
E.Religion

Answer: UC
Question: 164
Which scenario describes a potential risk and limitation of prompt engineering in the context of a generative AI
model?

A.Prompt engineering does not ensure that the model always produces consistent and deterministic outputs,
eliminating the need for validation.
B.Prompt engineering could expose the model to vulnerabilities such as prompt injection attacks.
C.Properly designed prompts reduce but do not eliminate the risk of data poisoning or model hijacking.
D.Prompt engineering does not ensure that the model will consistently generate highly reliable outputs when
working with real-world data.

Answer: B

Question: 165
A publishing company built a Retrieval Augmented Generation (RAG) based solution to give its users the ability to
interact with published content. New content is published daily. The company wants to provide a near real-time
experience to users.

Which steps in the RAG pipeline should the company implement by using offline batch processing to meet these
requirements? (Choose two.)

A.Generation of content embeddings

B.Generation of embeddings for user queries
C.Creation of the search index
D.Retrieval of relevant content
E.Response generation for the user

Answer: A

Question: 166
Which technique breaks a complex task into smaller subtasks that are sent sequentially to a large language model
(LLM)?

A.One-shot prompting
B.Prompt chaining
C.Tree of thoughts
D.Retrieval Augmented Generation (RAG)

Answer: B