/LLMDataHub

A quick guide (especially) for trending instruction finetuning datasets

MIT LicenseMIT

LLMDataHub: Awesome Datasets for LLM Training


🔥 Alignment Datasets • 💡 Domain-specific Datasets:atom: Pretraining Datasets 🖼️ Multimodal Datasets

GitHub last commit GitHub Repo stars

Introduction 📄

Large language models (LLMs), such as OpenAI's GPT series, Google's Bard, and Baidu's Wenxin Yiyan, are driving profound technological changes. Recently, with the emergence of open-source large model frameworks like LlaMa and ChatGLM, training an LLM is no longer the exclusive domain of resource-rich companies. Training LLMs by small organizations or individuals has become an important interest in the open-source community, with some notable works including Alpaca, Vicuna, and Luotuo. In addition to large model frameworks, large-scale and high-quality training corpora are also essential for training large language models. Currently, relevant open-source corpora in the community are still scattered. Therefore, the goal of this repository is to continuously collect high-quality training corpora for LLMs in the open-source community.

Training a chatbot LLM that can follow human instruction effectively requires access to high-quality datasets that cover a range of conversation domains and styles. In this repository, we provide a curated collection of datasets specifically designed for chatbot training, including links, size, language, usage, and a brief description of each dataset. Our goal is to make it easier for researchers and practitioners to identify and select the most relevant and useful datasets for their chatbot LLM training needs. Whether you're working on improving chatbot dialogue quality, response generation, or language understanding, this repository has something for you.

Contact 📬

If you want to contribute, you can contact:

Junhao Zhao 📧
Advised by Prof. Wanyun Cui

General Open Access Datasets for Alignment 🟢:

Type Tags 🏷️:

  • SFT: Supervised Finetune
    • Dialog: Each entry contains continuous conversations
    • Pairs: Each entry is an input-output pair
    • Context: Each entry has a context text and related QA pairs
  • PT: pretrain
  • CoT: Chain-of-Thought Finetune
  • RLHF: train reward model in Reinforcement Learning with Human Feedback

Datasets released in November 2023

Dataset name Used by Type Language Size Description ️
helpSteer / RLHF English 37k instances An RLHF dataset that is annotated by human with helpfulness, correctness, coherence, complexity and verbosity measures
no_robots / SFT English 10k instance High-quality human-created STF data, single turn.

Datasets released in September 2023

Dataset name Used by Type Language Size Description ️
Anthropic_
HH_Golden
ULMA SFT / RLHF English train 42.5k + test 2.3k Improved on the harmless dataset of Anthropic's Helpful and Harmless (HH) datasets. Using GPT4 to rewrite the original "chosen" answer. Compared with the original Harmless dataset, empirically this dataset improves the performance of RLHF, DPO or ULMA methods significantly on harmless metrics.

Datasets released in August 2023

Dataset name Used by Type Language Size Description ️
function_
calling_
extended
/ Pairs English
code
/ High quality human created dataset from enhance LM's API using ability.
AmericanStories / PT English / Vast sized corpus scanned from US Library of Congress.
dolma OLMo PT / 3T tokens A large diverse open-source corpus for LM pretraining.
Platypus Platypus2 Pairs English 25K A very high quality dataset for improving LM's STEM reasoning ability.
Puffin Redmond-Puffin
Series
Dialog English ~3k entries A dataset consists of conversations between real human and GPT-4,which features long context (over 1k tokens per conversation) and multi-turn dialogs.
tiny series / Pairs English / A series of short and concise codes or texts aim at improving LM's reasoning ability.
LongBench / Evaluation
Only
English
Chinese
17 tasks A benchmark for evaluate LLM's long context understanding capability.

Datasets released in July 2023

Dataset name Used by Type Language Size Description ️
orca-chat / Dialog English 198,463 entries An Orca-style dialog dataset aims at improving LM's long context conversational ability.
DialogStudio / Dialog Multilingual / A collection of diverse datasets aim at building conversational Chatbot.
chatbot_arena
_conversations
/ RLHF
Dialog
Multilingual 33k conversations Cleaned conversations with pairwise human preferences collected on Chatbot Arena.
WebGLM-qa WebGLm Pairs English 43.6k entries Dataset used by WebGLM, which is a QA system based on LLM and Internet. Each of the entry in this dataset comprise a question, a response and a reference. The response is grounded in the reference.
phi-1 phi-1 Dialog English / A dataset generated by using the method in Textbooks Are All You Need. It focuses on math and CS problems.
Linly-
pretraining-
dataset
Linly series PT Chinese 3.4GB Chinese pretraining dataset used by Linly series model, comprises ClueCorpusSmall, CSL news-crawl and etc.
FineGrainedRLHF / RLHF English ~5K examples A repo aims at develop a new framework to collect human feedbacks. Data collected is with the purpose to improve LLMs factual correctness, topic relevance and other abilities.
dolphin / Pairs English 4.5M entries An attempt to replicate Microsoft's Orca. Based on FLANv2.
openchat_
sharegpt4_
dataset
OpenChat Dialog English 6k dialogs A high quality dataset generated by using GPT-4 to complete refined ShareGPT prompts.

Datasets released in June 2023

Dataset name Used by Type Language Size Description ️
OpenOrca / Pairs English 4.5M completions A collection of augmented FLAN data. Generated by using method is Orca paper.
COIG-PC
COIG-Lite
/ Pairs Chinese / Enhanced version of COIG.
WizardLM_Orca orca_mini series Pairs English 55K entries Enhanced WizardLM data. Generated by using orca's method.
arxiv instruct datasets
math
CS
Physics
/ Pairs English 50K/
50K/
30K entries
dataset consists of question-answer pairs derived from ArXiv abstracts. Questions are generated using the t5-base model, while the answers are generated using the GPT-3.5-turbo model.
im-feeling-
curious
/ Pairs English 2595 entries Random questions and correspond facts generated by Google I'm feeling curious features.
ign_clean
_instruct
_dataset_500k
/ Pairs / 509K entries A large scale SFT dataset which is synthetically created from a subset of Ultrachat prompts. ⚠ lack of detailed datacard
WizardLM
evolve_instruct V2
WizardLM Dialog English 196k entries The latest version of Evolve Instruct dataset.
Dynosaur / Pairs English 800K entries The dataset generated by applying method in this paper. Highlight is generating high-quality data at low cost.
SlimPajama / PT Primarily
English
/ A cleaned and deduplicated version of RedPajama
LIMA dataset LIMA Pairs English 1k entries High quality SFT dataset used by LIMA: Less Is More for Alignment
TigerBot Series TigerBot PT
Pairs
Chinese
English
/ Datasets used to train the TigerBot, including pretraining data, STF data and some domain specific datasets like financial research reports.
TSI-v0 / Pairs English 30k examples
per task
A Multi-task instruction-tuning data recasted from 475 of the tasksource datasets. Similar to Flan dataset and Natural instruction.
NMBVC / PT Chinese / A large scale, continuously updating Chinese pretraining dataset.
StackOverflow
post
/ PT / 35GB Raw StackOverflow data in markdown format, for pretraining.

Datasets released before June 2023

Dataset name Used by Type Language Size Description ️
LaMini-Instruction / Pairs English 2.8M entries A dataset distilled from flan collection, p3 and self-instruction.
ultraChat / Dialog English 1.57M dialogs A large scale dialog dataset created by using two ChatGPT, one of which act as the user, another generates response.
ShareGPT_
Vicuna_unfiltered
Vicuna Pairs Multilingual 53K entries Cleaned ShareGPT dataset.
pku-saferlhf-dataset Beaver RLHF English 10K + 1M The first dataset of its kind and contains 10k instances with safety preferences.
RefGPT-Dataset
nonofficial link
RefGPT Pairs, Dialog Chinese ~50K entries A Chinese dialog dataset aims at improve the correctness of fact in LLMs (mitigate the hallucination of LLM).
Luotuo-QA-A
CoQA-Chinese
Luotuo project Context Chinese 127K QA pairs A dataset built upon translated CoQA. Augmented by using OpenAI API.
Wizard-LM-Chinese
instruct-evol
Luotuo project Pairs Chinese ~70K entries Chinese version WizardLM 70K. Answers are obtained by feed translated questions in OpenAI's GPT API and then get responses.
alpaca_chinese
dataset
/ Pairs Chinese / GPT-4 translated alpaca data includes some complement data (like Chinese poetry, application, etc.). Inspected by human.
Zhihu-KOL Open Assistant Pairs Chinese 1.5GB QA data on well-know Chinese Zhihu QA platform.
Alpaca-GPT-4_zh-cn / Pairs Chinese about 50K entries A Chinese Alpaca-style dataset, generated by GPT-4 originally in Chinese, not translated.
hh-rlhf
on Huggingface
Koala RLHF English 161k pairs
79.3MB
A pairwise dataset for training reward models in reinforcement learning for improving language models' harmlessness and helpfulness.
Panther-dataset_v1 Panther Pairs English 377 entries A dataset comes from the hh-rlhf. It rewrite hh-rlhf into the form of input-output pairs.
Baize Dataset Baize Dialog English 100K dialogs A dialog dataset generated by GPT-4 using self-talking. Questions and topics are collected from Quora, StackOverflow and some medical knowledge source.
h2ogpt-fortune2000
personalized
h2ogpt Pairs English 11363 entries A instruction finetune developed by h2oai, covered various topics.
SHP StableVicuna,
chat-opt,
, SteamSHP
RLHF English 385K entries An RLHF dataset different from previously mentioned ones, it use scores+timestamps to infer the users' preferences. Covers 18 domains, collected by Stanford.
ELI5 MiniLM series FT,
RLHF
English 270K entries Questions and Answers collected from Reddit, including score. Might be used for RLHF reward model training.
WizardLM
evol_instruct

V2
WizardLM Pairs English An instruction finetune dataset derived from Alpaca-52K, using the evolution method in this paper
MOSS SFT data MOSS Pairs,
Dialog
Chinese, English 1.1M entries A conversational dataset collected and developed by MOSS team. It has usefulness, loyalty and harmlessness labels for every data entries.
ShareGPT52K Koala, Stable LLM Pairs Multilingual 52K This dataset comprises conversations collected from ShareGPT, with a specific focus on customized creative conversation.
GPT-4all Dataset GPT-4all Pairs English,
Might have
a translated version
400k entries A combination of some subsets of OIG, P3 and Stackoverflow. Covers topics like general QA, customized creative questions.
COIG / Pairs Chinese,
code
200K entries A Chinese-based dataset. It contains domains like general purpose QA, Chinese exams, code. Its quality is checked by human annotators.
RedPajama-Data-1T RedPajama PT Primarily English 1.2T tokens
5TB
A fully open pretraining dataset follows the LLaMA's method.
OASST1 OpenAssistant Pairs,
Dialog
Multilingual
(English, Spanish, etc.)
66,497 conversation trees A large, human-written, human-annotated high quality conversation dataset. It aims at making LLM generates more natural response.
Alpaca-COT Phoenix Pairs,
Dialog,
CoT
English / A mixture a many dataset like classic Alpaca dataset, OIG, Guanaco and some CoT(Chain-of-Thought) datasets like FLAN-CoT. May be handy to use.
Bactrian-X / Pairs Multilingual
(52 languages)
67K entries per language A multilingual version of Alpaca and Dolly-15K.
databricks-dolly-15k
zh-cn Ver
Dolly2.0 Pairs English 15K+ entries A dataset of human-written prompts and responses, featuring tasks such as open-domain question-answering, brainstorming, summarization, and more.
AlpacaDataCleaned Some Alpaca/ LLaMA-like models Pairs English / Cleaned version of Alpaca, GPT_LLM and GPTeacher.
GPT-4-LLM Dataset Some Alpaca-like models Pairs,
RLHF
English,
Chinese
52K entries for English and Chinese respectively
9K entries unnatural-instruction
NOT the dataset used by GPT-4!! It is generated by GPT-4 and some other LLM for better Pairs and RLHF. It includes instruction data as well as comparison data in RLHF style.
GPTeacher / Pairs English 20k entries A dataset contains targets generated by GPT-4 and includes many of the same seed tasks as the Alpaca dataset, with the addition of some new tasks such as roleplay.
HC3 Koala RLHF English,
Chinese
24322 English
12853 Chinese
A multi-domain, human-vs-ChatGPT comparison dataset. Can be used for reward model training or ChatGPT detector training.
Alpaca data
Download
Alpaca, ChatGLM-finetune-LoRA, Koala Dialog,
Pairs
English 52K entries
21.4MB
A dataset generated by text-davinci-003 to improve language models' ability to follow human instruction.
OIG
OIG-small-chip2
Pythia-Chat-Base-7B, GPT-NeoXT-Chat-Base-20B, Koala Dialog,
Pairs
English,
code
44M entries A large conversational instruction dataset with medium and high quality subsets (OIG-small-chip2) for multi-task learning.
ChatAlpaca data / Dialog,
Pairs
English,
Chinese version coming soon
10k entries
39.5MB
A dataset aims to help researchers develop models for instruction-following in multi-turn conversations.
InstructionWild ColossalChat Pairs English, Chinese 10K enreues A Alpaca-style dataset, but with seed tasks comes from chatgpt screenshot.
Firefly(流萤) Firefly(流萤) Pairs Chinese 1.1M entries
1.17GB
A Chinese instruction-tuning dataset with 1.1 million human-written examples across 23 tasks, but no conversation.
BELLE
0.5M version
1M version
2M version
BELLE series, Chunhua (春华) Pairs Chinese 2.67B in total A Chinese instruction dataset similar to Alpaca data constructed by generating answers from seed tasks, but no conversation.
GuanacoDataset Guanaco Dialog,
Pairs
English,
Chinese,
Japanese
534,530 entries A multilingual instruction dataset for enhancing language models' capabilities in various linguistic tasks, such as natural language understanding and explicit content recognition.
OpenAI WebGPT WebGPT's reward model, Koala RLHF English 19,578 pairs Data set used in WebGPT paper. Used for training reward model in RLHF.
OpenAI
Summarization
Comparison
Koala RLHF English ~93K entries
420MB
A dataset of human feedback which helps training a reward model. The reward model was then used to train a summarization model to align with human preferences.
self-instruct / Pairs English 82K entries The dataset generated by using the well-known self-instruction method
unnatural-instructions / Pairs English 240,670 examples An early attempt to use powerful model (text-davinci-002) to generate data.
xP3 (and some variant) BLOOMZ, mT0 Pairs Multilingual,
code
79M entries
88GB
An instruction dataset for improving language models' generalization ability, similar to Natural Instruct.
Flan V2 / / English / A dataset compiles datasets from Flan 2021, P3, Super-Natural Instructions, along with dozens more datasets into one and formats them into a mix of zero-shot, few-shot and chain-of-thought templates
Natural Instruction
GitHub&Download
tk-instruct series Pairs,
evaluation
Multilingual / A benchmark with over 1,600 tasks with instruction and definition for evaluating and improving language models' multi-task generalization under natural language instruction.
CrossWOZ / Dialog English,
Chinese
6K dialogs The dataset introduced by this paper, mainly about tourism topic in Beijing, answers are generated automatically by rules.

Potential Overlaps ⚠️

We consider row items as subject.

OIG hh-rlhf xP3 natural instruct AlpacaDataCleaned GPT-4-LLM Alpaca-CoT
OIG / contains overlap overlap overlap overlap
hh-rlhf part of / overlap
xP3 overlap / overlap overlap
natural instruct overlap overlap / overlap
AlpacaDataCleaned overlap / overlap overlap
GPT-4-LLM overlap / overlap
Alpaca-CoT overlap overlap overlap overlap overlap overlap /

Open Datasets for Pretraining 🟢 :atom:

Dataset name Used by Type Language Size Description ️
proof-pile proof-GPT PT English
LaTeX
13GB A pretraining dataset which is similar to the pile but have LaTeX corpus to enhance LM's ability in proof.
peS2o / PT English 7.5GB A high quality academic paper dataset for pretraining.
StackOverflow
post
/ PT / 35GB Raw StackOverflow data in markdown format, for pretraining.
SlimPajama / PT Primarily
English
/ A cleaned and deduplicated version of RedPajama
NMBVC / PT Chinese / A large scale, continuously updating Chinese pretraining dataset.
falcon-refinedweb tiiuae/falcon series PT English / A refined subset of CommonCrawl.
CBook-150K / PT,
building dataset
Chinese 150K+ books A raw Chinese books dataset. Need some preprocess pipeline.
Common Crawl LLaMA (After some process) building datasets,
PT
/ / The most well-known raw dataset, rarely be used directly. One possible preprocess pipeline is CCNet
nlp_Chinese_Corpus / PT,
TF
Chinese / A Chinese pretrain corpus. Includes Wikipedia, Baidu Baike, Baidu QA, some forums QA and news corpus.
The Pile (V1) GLM (partly), LLaMA (partly), GPT-J, GPT-NeoX-20B, Cerebras-GPT 6.7B, OPT-175b PT Multilingual,
code
825GB A diverse open-source language modeling dataset consisting of 22 smaller, high-quality datasets that includes many domains and tasks.
C4
Huggingface dataset
TensorFlow dataset
Google T5 Series, LLaMA PT English 305GB A colossal, cleaned version of Common Crawl's web crawl corpus. Frequently be used.
ROOTS BLOOM PT Multilingual,
code
1.6TB A diverse open-source dataset consisting of sub-datasets like Wikipedia and StackExchange for language modeling.
PushshPairs reddit
paper
OPT-175b PT / / Raw reddit data, one possible processing pipeline in this paper
Gutenberg project LLaMA PT Multilingual / A book dataset, mostly novels. Not be preprocessed.
CLUECorpus / PT,
finetune,
evaluation
Chinese 100GB A Chinese pretraining Corpus sourced from Common Crawl.

Domain-specific Datasets 🟢 💡

Dataset name Used by Type Language Size Description ️
starcoderdata starcoder
series
PT code 783GB A large pretraining dataset for improving LM's coding ability.
code_
instructions
_120k_alpaca
/ Pairs English/code 121,959 entries code_instruction in instruction finetune format.
function-
invocations-25k
some MPT
variants
Pairs English code 25K entries A dataset aims at teaching AI models how to correctly invoke APIsGuru functions based on natural language prompts.
TheoremQA / Pairs English 800 A high quality STEM theorm QA dataset.
phi-1 phi-1 Dialog English / A dataset generated by using the method in Textbooks Are All You Need. It focuses on math and CS problems.
FinNLP FinGPT Raw data English,
Chinese
/ Open-source raw financial text data. Includes news, social media and etc.
PRM800K A variant of
GPT-4
Context English 800K entries A process supervision dataset for mathematical problems
MeChat data ⚠️use with care MeChat Dialog Chinese 355733 utterances A Chinese SFT dataset for training a mental healthcare chatbot.
ChatGPT-Jailbreak-Prompts
⚠️RISKY
/ / English 163KB file size Prompts for bypassing the safety regulation of ChatGPT. Can be use for probing the harmlessness of LLMs
awesome chinese
legal resources
LaWGPT / Chinese / A collection of Chinese legal data for LLM training.
Long Form / Pairs English 23.7K entries A dataset aims at improving the long text generation ability of LLM.
symbolic-instruction-tuning / Pairs English,
code
796 A dataset focuses on the 'symbolic' tasks: like SQL coding, mathematical computation, etc.
Safety Prompt / Evaluation only Chinese 100k entries Chinese safety prompts for evaluating and improving the safety of LLMs.
Tapir-Cleaned / Pairs English, 116k entries This is a revised version of the DAISLab dataset of PairsTT rules, which has been thoroughly cleaned, scored, and adjusted for the purpose of instruction-tuning
instructional_
codesearchnet_python
/ Pairs English &
Python
192MB This dataset is a template generated instructional Python datastet generated from an annotated version of the code-search-net dataset for the Open-Assistant project.
finance-alpaca / Pairs English 1.3K entries An Alpaca-style dataset but focus on financial topics

Multimodal Datasets for VLM

Dataset name Used by Type Language Size Description ️
ShareGPT4V / image-prompt-caption English 1.2M instances A set of GPT4-Vision-powered multi-modal captions data.
OBELICS idefics
series
image-document English 141M documents an open, massive, and curated collection of interleaved image-text web documents.
JourneyDB / image-prompt-caption English 4M instances A large scale dataset comprises QA, caption, and text prompting tasks, which is based on Midjourney images.
M3IT Ying-VLM instruction-image Multilingual 2.4M instances A dataset comprises 40 tasks with 400 human written instruction.
MIMIC-IT Otter instruction-image Multilingial 2.2M instances High quality multi-modal instructions-response pairs based on images and videos.
LLaVA Instruction LLaVA instruction-image English 158k samples A multimodal dataset generated upon COCO dataset by prompting GPT-4 to get instructions.

Private Datasets 🔴

Dataset name Used by Type Language Size Description ️
WebText(Reddit links) GPT-2 PT English / Data crawled from Reddit and filtered for GPT-2 pretraining.
MassiveText Gopher, Chinchilla PT 99% English, 1% other(including code)
WuDao(悟道) Corpora GLM PT Chinese 200GB A large scale Chinese corpus, Possible component originally open-sourced but not available now.