The Best

©Harshit Ahluwalia

LLMs
Part 1
LLMs Introduction
Cheatsheet
Basic Concepts How LLMs Work
Definition and Neural Networks & Tokenization
Overview Deep Learning
Attention Mechanism &
Importance and Training Data & Transformers
Applications Datasets
Language Modeling
Parameters & Model Objectives
Size

Popular LLM Training LLMs from Evaluation and

Architectures Scratch Metrics
GPT Data Collection & Perplexity
Preparation
BERT BLEU
Training Techniques
PaLM ROUGE
Fine-tuning & Transfer
LLaMA Learning GLUE

These parts are covered in this cheatsheet

 ©Harshit Ahluwalia

LLMs Introduction
Large Language Models (LLMs) are advanced deep learning
models designed to understand, generate, and manipulate
human language. They are typically based on neural network
architectures, such as transformers, which allow them to
process large amounts of text data and learn patterns in
human language.

These models are "large" due to their vast number of

parameters—often ranging from millions to hundreds of
billions—enabling them to capture complex language
structures and nuances.
BERT

LaMDa
GPT-4

LlaMa
Falcon 40B
 ©Harshit Ahluwalia

Importance and Applications

LLMs have revolutionized the field of natural language
processing (NLP) by significantly improving the accuracy and
fluency of language-related tasks. Their importance stems
from their ability to:

Understand and Generate Text: LLMs can perform various

tasks, including text completion, summarization, translation,
and question answering.

Adapt to Different Domains: By pre-training on vast datasets

and fine-tuning on specific tasks, LLMs can adapt to a wide
range of domains, from medical research to customer service.

Facilitate Human-Machine Interaction: They enable more

natural and intuitive interactions between humans and
machines, powering chatbots, virtual assistants, and other
conversational agents.

Application incude: Content Creation, Customer Support,

Data Analysis, Language Translation
 ©Harshit Ahluwalia

Basic Concepts and Terminology

Neural Networks and Deep Learning

Neural Networks are computational models inspired by

the human brain's structure, consisting of layers of
interconnected nodes (neurons) that process and learn
from data. Deep Learning is a subset of machine learning
that uses neural networks with many layers (deep
architectures) to model complex patterns in large
datasets.

LLMs, like GPT and BERT, are built on deep neural

networks, specifically transformer architectures, which
excel at processing sequences, such as text.
 ©Harshit Ahluwalia

Training Data and Datasets

Training Data is the foundational fuel for LLMs. It consists of

massive collections of text from diverse sources, such as
books, articles, websites, and scientific papers. These
datasets help the model learn the patterns, syntax, semantics,
and context of human language. Popular datasets for training
LLMs include:

Common Crawl: A dataset derived from web scraping,

providing a large, diverse text corpus.

Wikipedia and BooksCorpus: High-quality texts from

Wikipedia and books, often used to improve
understanding of structured language.

Specialized Datasets: Domain-specific datasets, such as

medical or legal texts, used to fine-tune LLMs for specific
tasks.
 ©Harshit Ahluwalia

Parameters and Model Size

Parameters are the components of the model that are learned

from the training data. They include weights and biases that
help the model make predictions. The Model Size is typically
measured by the number of parameters it contains. For
example:

Small Models: Contain millions to a few hundred million

parameters (e.g., GPT-2).

Medium to Large Models: Contain hundreds of millions to

a few billion parameters (e.g., BERT, GPT-3).

Very Large Models: Contain tens to hundreds of billions of

parameters (e.g., GPT-4, PaLM).

A larger number of parameters generally allows the model to

capture more complex patterns in data, but it also requires
more computational resources for training and inference.
 ©Harshit Ahluwalia

Pre-training and Fine-tuning

Pre-training is the initial phase where an LLM is trained

on a massive dataset in an unsupervised or self-
supervised manner to learn general language patterns.
During this phase, the model learns to predict the next
word in a sentence, fill in missing words, or generate
coherent text, thereby understanding grammar, facts, and
context.

Fine-tuning involves training the pre-trained model on a

smaller, task-specific dataset to adapt it to particular
applications. For example, an LLM pre-trained on general
text can be fine-tuned on a dataset of legal documents to
specialize in legal language and context. Fine-tuning
helps the model become more accurate and effective in
specific tasks, such as sentiment analysis, translation, or
question answering.
 ©Harshit Ahluwalia

How LLMs Work

Tokenization

Tokenization is the process of breaking down text into

smaller units, called tokens, which could be words,
subwords, or even individual characters. For example, the
sentence "ChatGPT is awesome!" might be tokenized into
["ChatGPT", "is", "awesome", "!"]. Tokenization is a crucial
first step in processing language because it converts
human-readable text into a format that the model can
understand and manipulate.

Word-level Tokenization: Splits text into individual words

but may struggle with out-of-vocabulary words (words not
seen during training).

Subword-level Tokenization (Byte-Pair Encoding,

WordPiece): Breaks down words into meaningful
subwords, allowing the model to handle rare or unseen
words more effectively.

Character-level Tokenization: Breaks text into individual

characters, useful for capturing the fine-grained structure
of languages but often less efficient for longer texts.
 ©Harshit Ahluwalia

Attention Mechanism and Transformers

The Attention Mechanism allows the model to focus on

specific parts of the input sequence when generating
output, effectively determining which words or tokens are
more relevant to the current task. It assigns different
weights to different words in a sentence, allowing the
model to capture context and dependencies more
accurately.

Transformers are neural network architectures built on the

attention mechanism. Unlike earlier models that
processed text sequentially, transformers can process all
words in a sequence simultaneously (in parallel), making
them highly efficient and capable of capturing long-range
dependencies in text. The key components of
transformers are:

Self-Attention

Multi-Head Attention

Feedforward Neural Networks

Positional Encoding
 ©Harshit Ahluwalia

Language Modeling Objectives

LLMs are trained using different language modeling

objectives depending on their architecture:

Masked Language Model (MLM): Used by models like

BERT. In MLM, some tokens in the input are randomly
masked, and the model is trained to predict these masked
tokens. This helps the model understand bidirectional
context—considering both the left and right context
around a word.

Causal Language Model (CLM): Used by models like

GPT. CLM predicts the next word in a sequence given all
the previous words. It is unidirectional, meaning it only
considers the left context (past tokens) to predict the next
token, making it useful for generative tasks where the
output is generated sequentially, like text completion.
 ©Harshit Ahluwalia

Inference and Sampling Methods

Inference is the process of generating text or performing a task using a

trained LLM. During inference, the model takes an input prompt and
produces a relevant output based on the patterns and information it has
learned during training.

Sampling Methods are techniques used to decide which word or

token to generate next during inference. Common methods include:

Greedy Search: Selects the token with the highest probability at

each step. It is fast but can produce repetitive or less diverse
outputs.

Beam Search: Evaluates multiple possible outputs at each step

and keeps the most likely sequences. It balances between
exploration (trying different options) and exploitation (choosing the
best path), often producing more coherent results.

Top-k Sampling: Chooses from the top-k most probable tokens at

each step, allowing for more diverse outputs by adding
randomness.

Nucleus Sampling (Top-p Sampling): Chooses from the smallest

set of tokens whose cumulative probability exceeds a certain
threshold (p), balancing between diversity and coherence.
 ©Harshit Ahluwalia

Popular LLM Architectures

GPT (Generative Pre-trained Transformer)

GPT (Generative Pre-trained Transformer) is an LLM developed by

OpenAI, known for its generative capabilities. GPT models are
based on a Causal Language Model (CLM), where the goal is to
predict the next word in a sequence, given all the previous words.
The architecture is unidirectional, focusing on the left-to-right
context, making GPT particularly powerful for tasks that involve
generating coherent and contextually relevant text, such as text
completion, storytelling, and creative writing.

Key characteristics of GPT:

Transformer Decoder Architecture: Uses only the decoder part of

the transformer, optimized for generating text.

Pre-training and Fine-tuning: Pre-trained on a large corpus of

diverse text and fine-tuned on specific tasks to adapt to different
applications.

Generative Focus: Excellent at tasks where generating new text is

essential (e.g., dialogue generation, content creation).

Versions: Includes GPT, GPT-2, GPT-3, and GPT-4, with each new
version incorporating larger model sizes and improved
performance.
 ©Harshit Ahluwalia

Bidirectional Encoder
BERT Representations from Transformers

BERT (Bidirectional Encoder Representations from Transformers)

is a model developed by Google that introduced a new way of
understanding language context by considering both the left and
right sides of a word simultaneously, hence "bidirectional." BERT is
designed for Masked Language Modeling (MLM), where random
words in a sentence are masked, and the model is trained to
predict these words based on their context. This makes BERT
highly effective for tasks that require deep understanding and
contextual awareness, such as sentiment analysis, named entity
recognition, and question answering.

Key characteristics of BERT:

Transformer Encoder Architecture: Utilizes the encoder part of the

transformer to focus on understanding the meaning of the input
text.

Bidirectional Context: Reads text from both directions to capture full

context, enhancing comprehension for NLP tasks.

Pre-training on Large Datasets: Pre-trained on vast amounts of text

from sources like Wikipedia, BooksCorpus, etc.

Fine-tuning for Specific Tasks: Can be fine-tuned for specific NLP

tasks using small, task-specific datasets.

Variants: Includes BERT, RoBERTa, DistilBERT, and others that

build on or modify the original BERT architecture for different
needs.
 ©Harshit Ahluwalia

Text-to-Text Transfer
T5 Transformer

T5 (Text-to-Text Transfer Transformer), developed by Google, is an

innovative approach to NLP tasks by framing every problem as a
text-to-text task. This means both the input and output are always
text strings, regardless of the task (e.g., translation, summarization,
or sentiment analysis). T5 is trained on a diverse set of NLP tasks,
which allows it to generalize well across different applications.

Key characteristics of T5:

Unified Text-to-Text Framework: Handles all NLP tasks in a

consistent format by treating every input and output as text.

Transformer Architecture: Uses both the encoder and decoder

parts of the transformer, providing flexibility in understanding and
generating text.

Multi-task Learning: Trained on a large collection of different NLP

tasks simultaneously, improving generalization and adaptability.

Pre-training and Fine-tuning: Like other LLMs, T5 is pre-trained

on massive datasets and can be fine-tuned for specific applications.

Scalability: Available in various sizes (small, base, large, XL, XXL)

to balance between computational efficiency and performance.
 ©Harshit Ahluwalia

LLaMA, PaLM, and Others

LLaMA (Large Language Model Meta AI): A family of foundational

language models by Meta (formerly Facebook), optimized for
efficiency and performance. LLaMA is designed to democratize
access to LLM capabilities by providing smaller, more efficient
models that are still competitive in performance with much larger
counterparts. It aims to provide robust language understanding and
generation with reduced computational requirements.

PaLM (Pathways Language Model): Developed by Google, PaLM

is a large, dense language model that leverages the Pathways
system, designed to enable a single model to be trained on many
tasks, with better efficiency and resource utilization. PaLM is highly
scalable, designed to learn from vast amounts of data across
multiple modalities, and is geared toward advanced NLP tasks like
reasoning, commonsense understanding, and creative text
generation.
Other Notable Models

XLNet: Combines ideas from both BERT and autoregressive

models to capture bidirectional context while addressing limitations
of BERT’s masked language modeling.

Megatron-Turing NLG: A large-scale, generative model jointly

developed by NVIDIA and Microsoft, designed for natural language
generation tasks.

GLaM (Gated Language Model): Developed by Google, GLaM

uses a mixture-of-experts approach, allowing it to dynamically
activate different parts of the model for different tasks, making it
more efficient in resource use.
Moreover,
we are offering a
Free Certification
on LLMs, check the link
in the description

@Harshit Ahluwalia