Skip to content

FreedomIntelligence/InstructionZoo

BranchesTags

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

63 Commits
README.md
README.md
 
 

Repository files navigation

InstructionZoo

A collection of open-source Instruction-tuning dataset to train chat-based LLMs (ChatGPT,LLaMA,Alpaca).

This is an on-going project. We will soon add tags to classify the following datasets and continuously update our collection.

Table of Contents

The template

## [owner/project-name](https://github.com/link/to/project)

* Size:
* Language:
* Summary:
* Generation Method:
* Paper:
* HuggingFace: (if applicable)
* Demo: (if applicable)
* License:

The English Instruction Datasets

tatsu-lab/Alpaca

gururise/Cleaned Alpaca

PhoebusSi/Alpaca-COT

QingyiSi/Alpaca-CoT

  • Empty for now. Soon to update.

orhonovich/unnatural-instructions

  • Size: 240,000 instructions
  • Language: EN
  • Summary: Unnatural Instructions consist of a core dataset of 68,478 instruction-input-output triplets, and a full dataset.
  • Generateion Method:
    • Step 1 (Core Dataset Generation): Collect 64,000 examples by prompting a language model with three seed examples of instructions and eliciting a fourth, following a strict instruction-input-output format.
    • Step 2 (Template Expansion): Prompt a language model to reformulate the tasks in the core dataset, and collect two alternative formulations for each generated task
  • Paper: Unnatural Instructions: Tuning Language Models with (Almost) No Human Labor
  • License:

bigscience/PromptSource

bigscience/P3

allenai/natural-instructions

allenai/super-natural-instructions

google-research/FLAN 2021

  • Size: 62 tasks
  • Language: EN
  • Summary: FLAN 2021 aggregates 62 text datasets on Tensorflow Datasets into a single mixture. It is currently not public.
  • Generateion Method: Map exist datasets into Instruction Schema.
  • Paper: Finetuned Language Models Are Zero-Shot Learners
  • License:

google-research/FLAN 2022 Collection

tasksource-instruct

LianjiaTech/BELLE 1.5M

LianjiaTech/BELLE 10M

XueFuzhao/InstructionWild

  • Size: 479 seed instructions, 52,191 Chinese instructions, 52,191 English instructions
  • Language: CH, EN
  • Summary: InstructionWild use the same format as Alpaca for fast and easy usage. Its instructions have no input field.
  • Generateion Method:
    • Pick 429 instructions over 700 noisy instructions from Twitter
    • Use a similar method as Alpaca for generating the resulting instructions.
  • License:

ExMix

UnifiedSKG

MetaICL

openai/InstructionGPT

facebookresearch/metasqe/OPT-IML

  • Size: 1,667 tasks, 3,128 instructions
  • Language: EN
  • Summary: OPT-IML dataset expands the Super-Natural-Instructions benchmark with the task collections from multiple existing work on instruction-tuning, cross-task transfer studies, and area-specific task consolidation.
  • Generation Method:
    • Benchmarks included in OPT-IML are Super-Natural-Instructions, PromptSource, CrossFit, FLAN, ExMix, T5, UnifiedSKG, and Reasoning. Authors only kept partial tasks from CrossFit, ExMix and T5 due to the significant overlap.
    • To organize the Instruction schema, authors broadly classify the instructions in these benchmarks into two categories, dataset-level and instance-level.
  • Paper: OPT-IML: Scaling Language Model Instruction Meta Learning through the Lens of Generalization
  • License:

THUDM/GLM-130B

laion/OIG

baize/baize-chatbot

lightaime/camel

thunlp/UltraChat

  • Size: 657K instructions
  • Language: EN
  • Summary: UltraChat is a multi-round dialogue dataset powered by Turbo APIs, composed of three sectors, namely Questions about the World, Writing and Creation, and Assistance on Existent Materials.
  • Generation Method:
    • Two separate ChatGPT Turbo APIs are adopted in generation, where one plays the role of the user to generate queries and the other generates the response.
    • We instruct the user model with carefully designed prompts to mimic human user behavior and call the two APIs iteratively.
  • HuggingFace: https://huggingface.co/datasets/stingning/ultrachat
  • License:

databrickslabs/doll

  • Size: 7 tasks, 15,000 instructions
  • Language: EN
  • Summary: Dolly is a human-generated corpus, whose categories are Creative Writing, Closed QA, Open QA, Summarization, Information Extraction, Classification and Brainstorming.
  • Generation Method:
    • Databricks employees were invited to create prompt / response pairs in each of eight different instruction categories.
    • For instruction categories that require an annotator to consult a reference text, contributors selected passages from Wikipedia for particular subsets of instruction categories.
  • HuggingFace: https://huggingface.co/datasets/databricks/databricks-dolly-15k
  • License:

Instruction-Tuning-with-GPT-4/GPT-4-LLM

ShareGPT

stanfordnlp/SHP

Anthropic/hh-rlhf

HuggingFaceH4/stack-exchange-preferences

Hellp-SimpleAI/HC3

f/awesome-chatgpt-prompts

  • Empty for now. Soon to update.

The Chinese Instruction Datasets

FlagOpen/FlagInstruct

  • Size: 2K tasks, 191,191 instructions in total
  • Language: CH
  • Summary: Chinese Open Instruction Generalist (COIG) is a Chinese instruction dataset consisting of 4 sub-tasks.
  • Generateion Method:
    • Task 1: Translated Instructions (67,798)
      • Translate the following datasets into Chinese: 1,616 task descriptions in Super-Natural-Instruct v2 along with a single instance for each of them; 175 seed tasks in Self-instruct; 66,007 instructions from Unnatural Instructions.
    • Task 2: Exam Instructions (63,532)
      • Exams include The Chinese National College Entrance Examination (高考), Middle School Entrance Examinations (中考), and Civil Servant Examination (公务员考试).
      • Turn them into Chain-of-Thought (CoT) corpus by extracting six informative elements from original exam questions, including instruction, question context, question, answer, answer analysis, and coarse-grained subject.
    • Task 3: Human Value Alignment Instructions (34,471)
      • Select a set of samples that present shared human values in the Chinese-speaking world, and get 50 seed instructions and 3k resulting instructions.
      • Some additional sets of samples that present regional-culture or country-specific human values are also added.
    • Task 4: Counterfactural Correction Multi-round Chat (13,653)
      • The aim is to alleviate and resolve the pain points of hallucination and factual inconsistency in current LLMs.
      • Based on CN-DBpedia knowledge graph dataset, CCMC has ~13,000 dialogues with an average of 5 rounds per dialogue, resulting in ~65,000 rounds of chat.
    • Leetcode Instructions (11,737)
      • 2,589 programming questions from Leetcode.
  • Paper: Chinese Open Instruction Generalist: A Preliminary Release
  • HuggingFace: https://huggingface.co/datasets/BAAI/COIG
  • License: MIT License

CLUEbenchmark/pCLUE

ydli-ai/CSL

  • Size: 4 tasks, 396,209 instructions
  • Language: CH
  • Summary: CSL is a large-scale Chinese scientific literature dataset.
  • Generation Method:
    • Obtain the paper’s meta-information from the National Engineering Research Center for Science and Technology Resources Sharing Service (NSTR) dated from 2010 to 2020.
    • Label papers with categories and disciplines, with the assistance of volunteers.
    • The data format in CSL is <T,A,K,c,d>, where T is the title, A is the abstract, K is a list of keywords, c is the category label and d is the discipline label.
  • Paper: CSL: A Large-scale Chinese Scientific Literature Dataset
  • License:

YeungNLP/Firefly

TsinghuaAI/CUGE

ydli-ai/Chinese-ChatLLaMA

  • Language: Multilingual
  • License:

ZeroPrompt

PlexPt/awesome-chatgpt-prompts-zh

  • Empty for now. Soon to update.

Chinese Alpaca

carbonz0/alpaca-chinese-dataset

  • Size: 20,456 instructions
  • Language: CH
  • Generateion Method: Translate Alpaca into Chinese by machine and then clean.

hikariming/alpaca_chinese_dataset

  • Size: 19,442 instructions
  • Language: CH
  • Generateion Method: Translate Alpaca into Chinese by ChatGPT, and check them by humans

ymcui/Chinese-LLaMA-Alpaca

  • Size: 51,458 instructions
  • Language: CH
  • Generateion Method: Translate Alpaca into Chinese by ChatGPT, and discard some of them.

LC1332/Chinese-alpaca-lora

  • Size: 51,672 instructions
  • Language: CH
  • Generateion Method: Translate Stanford Alpaca dataset into Chinese by ChatGPT.

A-baoYang/alpaca-7b-chinese

  • Size: 20,465 instructions
  • Language: TC
  • Generateion Method: Translate Stanford Alpaca dataset into traditional Chinese using OpenCC.

A-baoYang/alpaca-7b-chinese

  • Size: 124,469 instructions
  • Language: EN, TC
  • Generateion Method: Combine the English instruction/input and traditional Chinese output by ChatGPT.

ntunlplab/traditional-chinese-alpaca

  • Size: 52,002 instructions
  • Language: EN, TC
  • Generateion Method: A Traditional-Chinese version of the Alpaca dataset, whose instruction part is left as English.

ntunlplab/traditional-chinese-alpaca

  • Size: 52,002 instructions
  • Language: EN, TC
  • Generateion Method: An Traditional-Chinese version of the Alpaca dataset, where there are English and traditional Chinese versions of one single instruction.

The Miltilingual Instruction Datasets

bigscience/xP3

JosephusCheung/GuanacoDataset

  • Size: 380,835 instructions in total
  • Language: CH, DE, EN, JA, TC
  • Summary: Guanaco dataset builds upon the 175 tasks from Alpaca, containing 3 versions with different sizes and methods.
  • Generateion Method:
    • Original Version (48967): Rewrite 175 Alpaca seed tasks in different languages, and add new tasks specifically designed for English grammar analysis, natural language understanding, cross-lingual self-awareness, and explicit content recognition.
    • Mixed Version (279644): The original 175 tasks were translated into 4 versions and regenerated independently, excluding Deutsch.
    • MIni Version (52224): 52K instrucrion dataset, which is included in the Mixed Version.
  • HuggingFace: https://huggingface.co/datasets/JosephusCheung/GuanacoDataset/tree/main
  • License:

JosephusCheung/GuanacoDataset QA

  • Size: 205,999 instructions in total
  • Language: CH, DE, EN, JA
  • Summary: The Paper/General-QA dataset is a collection of questions and answers constructed for AI-generated papers or general texts in 4 languages. The purpose of this dataset is to generate paragraph-level answers to questions posed about lengthy documents such as PDFs.
  • Generateion Method:
    • The question dataset contains 106,707 questions, and the answer dataset contains 99,292 answers.
    • Similar questions are combined to form a tree-like structure, and graph theory algorithms are used to process user questions, content summaries, and contextual logic.
  • HuggingFace: https://huggingface.co/datasets/JosephusCheung/GuanacoDataset/tree/main/additional
  • License:

The Code Instruction Datasets

sahil280114/codealpaca

  • Size: 20,023 instructions
  • Language: EN
  • Summary:
  • Generateion Method: Self-instuct with prompts to focus on code generation/edting/optimization tasks, using text-davinci-003.
  • HuggingFace:
  • License:

About

No description, website, or topics provided.

Resources

Readme
Activity
Custom properties

Stars

282 stars

Watchers

11 watching

Forks

24 forks
Report repository

Releases

No releases published

Packages

No packages published

Contributors 4

  •  
  •  
  •  
  •