awesome-chatgpt-dataset

Unlock the Power of LLM: Explore These Datasets to Train Your Own ChatGPT!

Select your own mixed dataset

git clone https://github.com/voidful/awesome-chatgpt-dataset.git
cd awesome-chatgpt-dataset/mixed/dataset

pick whatever dataset you want to use, then merge and upload:

python preprocess.py your_dataset_name_to_HuggingFaceHub

Dataset Detail

Sorted by dataset size (small → large). Items with unknown size appear at the end.

Dataset Name	Size	Languages	Source	License
TheoremQA	1K	English	We annotated 800 QA pairs covering 350+ theorems spanning across Math, EE&CS, Physics and Finance.	mit
LIMA	1K	English	LIMA: Less Is More for Alignment.	cc-by-nc-sa-4.0
WildGuardMix	1.7K	English	Safety training mixture with vanilla/adversarial prompts and multi-annotator labels.	odc-by
Berkeley Function Calling Leaderboard (BFCL)	2K	English + code	Function/tool-calling eval covering parallel/multi-call scenarios across languages.	-
ruozhiba_s2t	2K	Traditional Chinese	Ruozhiba (弱智吧) logic traps and witty humorous Q&A converted to Traditional Chinese.	Apache-2.0
im-feeling-curious	3K	English	Extract from Google’s “I’m Feeling Curious” facts.	-
Puffin	3K	English	Exactly 3,000 multi-turn examples; each response via GPT‑4.	apache-2.0
cc_sbu_align	4K	English	MiniGPT‑4 alignment data (image–text).	bsd-3-clause
QA-Feedback	4K	English	Re‑constructed ASQA with human feedback.	-
SLF5K	5K	English	Summarization with Language Feedback (5K unique samples).	apache-2.0
blended_skill_talk	7K	English	7k conversations blending personality, empathy, and knowledge.	-
GSM‑IC	8K	English	Grade‑School Math with Irrelevant Context (distractor sentences).	-
ChatAlpaca‑10K	10K	English	10,000 multi‑turn conversations (Alpaca‑based).	apache-2.0
PKU‑SafeRLHF‑10K	10K	English	First‑round Safe‑RLHF data with safety preferences.	-
deita-10k-v0	10K	English	Data-Efficient Instruction Tuning for Large Language Models.	MIT
Dolly‑15K	15K	English	15k instruction records crowdsourced by Databricks.	cc-by-3.0
Capybara	16K	English	High-quality multi-turn conversation and instruction dataset.	Apache-2.0
Bespoke-Stratos-17k	17K	English	Distilled reasoning and problem-solving dataset for training deep-reasoning and chain-of-thought capabilities.	MIT
WebGPT (comparisons)	20K	English	Human preference comparisons for WebGPT reward modeling.	-
CodeAlpaca‑20K	20K	English	20,022 instruction–code pairs for code generation.	-
HelpSteer2	21K	English	Open-source helpfulness data for reward models and preference learning.	cc-by-4.0
openapi-function-invocations‑25k	25K	English	Synthetic + extracted OpenAPI function-call traces.	mit
LongForm	28K	English	Reverse‑instruction long‑text generation dataset.	mit
Chatbot Arena Conversations	33K	English	33K cleaned Arena chats with pairwise preferences.	-
HC3	37K	English, Chinese	37,175 instructions with human vs LLM answers.	-
Anthropic HH Golden	45K	English	Helpful & Harmless preference data; golden subset.	-
COIG-CQIA	45K	Chinese	Chinese Open Instruction Generalist - Quality is All You Need, high quality human-written Chinese instruction dataset from diverse internet sources.	Apache-2.0
Mol‑Instructions	48K	English	Biomolecular instruction dataset for LLMs.	cc-by-4.0
RefGPT	50K	English, Chinese	Cost‑effective pipeline to generate multi‑turn Q&A with references.	-
arxiv‑math‑instruct‑50k	50K	English	QA pairs derived from arXiv math abstracts.	-
arxiv‑math‑instruct‑50k (ArtifactAI)	51K	English	T5‑generated questions; GPT‑3.5 answers.	-
Traditional Chinese Alpaca	52K	Traditional Chinese	Alpaca translated by ChatGPT API.	apache-2.0
Cabrita Dataset	52K	Portuguese	Alpaca translated to Portuguese.	-
Japanese Alpaca	52K	Japanese	Alpaca translated by ChatGPT API.	cc-by-nc-4.0; OpenAI terms
Alpaca Dataset	52K	English	175 seed instructions completed by OpenAI.	cc-by-nc-4.0; OpenAI terms
Alpaca Data Cleaned	52K	English	Cleaned Alpaca 52K.	-
Alpaca GPT‑4 Data	52K	English	Same prompts, GPT‑4 completions.	-
Alpaca GPT‑4 Chinese	52K	Chinese	GPT‑4 completions for Chinese prompts.	-
xLAM Function Calling 60K	60K	English	Structured tool-calling data for executable agents.	apache-2.0
UltraFeedback	64K	English	Large-scale, fine-grained preference dataset annotated across instruction-following, truthfulness, honesty, and helpfulness.	MIT
Dynosaur	66K	English	Dynamic growth paradigm for instruction curation.	apache-2.0
Finance	69K	English	68,912 finance‑related instructions.	-
WizardLM evol	70K	English	Evolutionary instruction tuning data (WizardLM).	-
Vicuna Dataset	75K	English	~100k ShareGPT conversations (curated).	-
Magicoder-OSS-Instruct-75K	75K	English, Code	Synthetic coding instruction dataset generated by prompting LLMs with seeds from real open-source code snippets.	MIT
InstructionTranslation	80K	Multi-lingual	M2M‑12B translated instructions (≤512 tokens).	mit
Self‑Instruct	82K	English	52K seed instructions; 82K I/O pairs.	-
OASST1	89K	Multi-lingual	Human‑generated assistant conversations (35 languages).	apache-2.0
HH‑RLHF	91K	English	Helpful/harmless RLHF pairs.	mit
Guanaco Dataset	98K	En, Zh‑CN, Zh‑HK/TW, Ja	175 Alpaca tasks across languages.	gpl-3.0
FineTome-100k	100K	English	High-quality instruction dataset curated and filtered using the FineWeb-Edu classifier for maximum educational value.	Apache-2.0
ChatDoctor	100K	English	Medical question-answering dataset based on real doctor-patient consultations from HealthCareMagic.	-
InstructionWild	104K	English, Chinese	Seeded 429 instructions; ~52K generated.	research-only; OpenAI terms
CAMEL Dataset	107K	English	Multi‑role, topic‑diverse instruction dialogues.	-
glaive-function-calling-v2	113K	English	Function calling and tool-use dataset with detailed JSON schemas, API call generation, and conversational follow-ups.	Apache-2.0
OpenThoughts-114k	114K	English	Open synthetic reasoning dataset with detailed step-by-step thinking traces covering math, science, and coding.	MIT
TAPIR‑Cleaned	117K	English	Cleaned IFTTT rule dataset for instruction tuning.	cc-by-nc-4.0
ToolBench	120K	English	An instruction-tuning dataset for general tool-use and multi-step API execution across 16,000+ real-world REST APIs.	Apache-2.0
OASST2 (final)	135K	Multi-lingual	Open Assistant Conversations Release 2 (train+val).	apache-2.0
glaive-code-assistant	136K	English	High quality coding instructions and programming assistant conversations.	-
WizardLM Evol‑Instruct V2	143K	English	143K mixture‑evolved data.	-
LLaVA Visual Instruct 150K	150K	English	GPT‑generated multimodal instruction pairs.	cc-by-nc-4.0
CodeFeedback-Filtered-Instruction	156K	English, Code	Filtered, diverse coding instruction dataset collected from Magicoder and ShareGPT-Python for code LLMs.	Apache-2.0
ProsocialDialog	166K	English	165,681 prosocial instructions and feedback.	-
M2Lingual	175K	Multi-lingual	Multilingual mixed‑modal (code+text) chat/instruct SFT.	-
Sujet-Finance-Instruct-177k	177K	English	Comprehensive financial instruction dataset spanning analysis, summarization, and domain calculations.	Apache-2.0
Nectar	183K	English	183K prompts each with 7 model responses ranked by GPT-4 for preference and reward model training.	Apache-2.0
COIG	191K	Chinese	Chinese Open Instruction Generalist.	apache-2.0
orca‑chat	198K	English	Cleaned, pruned conversation‑style Orca subset.	-
chatdoctor-200k-stripped	200K	English	Cleaned medical consultation dialogue dataset between patients and physicians.	-
OpenR1‑Math‑220k	220K	English	DeepSeek‑R1 distilled math traces (verified).	apache-2.0
Unnatural Instructions	241K	English	Large creative/diverse instruction corpus.	mit
WildJailbreak	262K	English	Synthetic jailbreak and benign contrastive prompts.	odc-by
Magpie-Llama-3.1-Pro-300K-Filtered	300K	Multi-lingual	Multi-turn high-quality synthesis dataset generated by prompting aligned Llama 3.1 with nothing.	CC-BY-NC-4.0
tulu-v2-sft-mixture	326K	English	Supervised fine-tuning mixture used to train Ai2 Tülu 2 models.	ODC-By
SHP	358K	English	385K Reddit preference pairs across 18 topics.	reddit – revocable, non‑exclusive
Dromedary	361K	English	Dromedary‑Verbose‑Clone synthetic instructions.	cc-by-nc-4.0
MetaMathQA	395K	English	Mathematical reasoning dataset augmented from GSM8K and MATH via bootstrapping and re-writing.	Apache-2.0
UltraChat	404K	English	Dual‑API generation (user vs assistant) for quality control.	cc-by-nc-4.0
IGN Clean Instruct 500K	509K	English	~508k Ultrachat‑sourced, high‑quality instructions.	apache-2.0
SlimOrca	518K	English	Curated, high-quality subset of OpenOrca rich in step-by-step GPT-4 reasoning traces.	MIT
ELI5	559K	English	Long‑form community Q&A (“Explain Like I’m Five”).	-
Xerxes-Instruct-700K	700K	Multi-lingual	Multilingual instruction tuning dataset covering diverse conversational and downstream tasks.	CC-BY-4.0
GPT4All	806K	Multi-lingual	LAION OIG + StackOverflow + P3 prompts; OpenAI outputs.	-
NuminaMath-CoT	860K	English	Math Olympiad and competitive exam problems with step-by-step Chain-of-Thought solutions.	Apache-2.0
Instruct	889K	English	888,969 English instructions (augmented).	mit
tulu-3-sft-mixture	939K	English	Ai2 Tülu 3 supervised fine-tuning dataset mixture covering general chat, math, coding, and reasoning.	ODC-By
OpenHermes-2.5	1M	English, Code	1M high-quality GPT-4 distilled instruction dataset with substantial code and multi-turn conversations.	Apache-2.0
WildChat	1.0M	English	In‑the‑wild user–LLM chat dataset (license updated).	odc-by
lmsys-chat-1m	1M	Multi-lingual	1M real-world human interactions with 25 LLMs gathered from LMSYS Chatbot Arena.	LMSYS Terms
MOSS	1M	Chinese	GPT‑3.5‑turbo generated Chinese SFT data.	apache-2.0 + agpl-3.0
smolTalk	1.1M	English	Ultra‑compact multi‑turn chat for small‑scale SFT.	apache-2.0
ShareGPT4V	1.2M	English	High-quality image-text captions and visual multi-turn instruction conversations generated by GPT-4-Vision.	CC-BY-NC-4.0
Open‑PerfectBlend	1.42M	English	Diverse, deduped chat blend for general SFT.	apache-2.0
Firefly	1.6M	Chinese	1.6M+ Chinese instructions across 23 NLP tasks.	-
The Tome	1.75M	English	Large cleaned instruction dataset curated by Arcee.	mit
NaturalReasoning	2.8M	English	2.8M challenging reasoning questions (decontaminated).	cc-by-nc-4.0
LaMini‑Instruction	3.0M	English	~2.58M–3M instruction–response pairs (GPT‑3.5).	cc-by-nc-4.0
OpenOrca (full)	3.0M	English	GPT‑4/3.5 augmented FLAN collection.	-
WildChat‑4.8M (nontoxic subset)	3.20M	English	Nontoxic filtered split of WildChat 4.8M.	odc-by
Natural Instructions	5M	Multi-lingual	5,040,134 instructions collected from diverse NLP tasks	-
OpenMathInstruct-2	5M	English	Math reasoning dataset with 5M verified solutions generated by Llama-3.1-405B-Instruct.	CC-BY-4.0
Infinity‑Instruct	8.9M	Multi-lingual	7.4M base + ~1.5M chat instruction data.	cc-by-sa-4.0
BELLE‑10M	10M	Chinese	Multi‑type Chinese instructions across domains.	research-only; OpenAI terms
OIG‑43M	43M	Multi-lingual	LAION + Together + OntoCord composite instruction pool.	-
xP3	79M	Multi-lingual	78,883,588 instructions from prompted datasets across 46 languages & 16 tasks.	-

Unknown / mixed-size (kept for completeness; format consistent with original)

Dataset Name	Size	Languages	Source	License
TableLLM-SFT	-	English	Benchmark & SFT instruction tuning dataset tailored for table understanding, reasoning, and operations.	-
CodeParrot	-	python	180GB Python files (<1MB each), 20M+ files.	-
Alpaca‑CoT	-	Multi-lingual	Instruction data with chain‑of‑thought traces.	odc-by
stack-exchange-paired	-	English	StackExchange Q&A pairs for preference modeling.	cc-by-sa-4.0
LangChainDatasets	-	English	Community datasets to evaluate chains & agents.	-
ParlAI	-	English	Dialog research platform with many tasks/datasets.	-
GPTeacher	-	English	Instruction datasets consolidated for general SFT.	-
Wizard‑LM Chinese Evol	-	Chinese	Chinese evol‑instruct corpus.	-
MultiWOZ	-	English	Multi‑domain Wizard‑of‑Oz dialog dataset.	-
ToolACE	-	English	Multi‑tool calling SFT (functions, API JSON, tool plans).	-
UltraFeedback (cleaned binarized)	-	English	UltraFeedback preferences cleaned & binarized.	cc-by-nc-4.0