Skip to content

Datasets

Hugging Face datasets with plain-language descriptions: what's inside and what models to train on them. 4,788 in the catalog so far, with new ones added every hour.

Data type

For which models

Find a dataset for your goal

Answered by Gemini 4 · the chat is kept only in this tab

DatasetHuggingFaceBio/FineDNATop pick

Carbon-X training data for DNA reading

Three training stages for Carbon-X: mid-training, SFT, and RL on DNA texts, dialogues, and prompts.

  • Text generation
  • English
13.7K 70100M–1B rows
Datasetkotekjedi/stolen-thoughtsTop pick

Recovered hidden reasoning of 27 closed LLMs

Dataset for research on extracting hidden reasoning chains from proprietary APIs on complex tasks.

  • Text generation
  • English
13.6K 810K–100K rows🔒 on request
DatasetSWE-bench/SWE-smith-trajectories

SWE-smith trajectories for coding agent

5,017 SWE-agent + Claude 3.7 Sonnet trajectories used for fine-tuning Qwen 2.5 Coder Instruct.

  • Text generation
  • English
13.6K 8210K–100K rows
DatasetRUC-NLPIR/FlashRAG_datasets

36 RAG benchmarks in the FlashRAG suite

Preprocessed datasets for RAG research: 36 benchmarks and FlashRAG resources.

  • Question answering
  • Summarization
  • English
13.5K 961M–10M rows
DatasetTIGER-Lab/MathInstruct

Math instructions with CoT and PoT

Math problems with reasoning chains and programs for training instruction models.

  • Text generation
  • English
13.4K 321100K–1M rows
DatasetHAERAE-HUB/KMMLU

Korean KMMLU benchmark

35,030 expert multiple-choice questions across 45 subjects in Korean.

  • Multiple choice
  • Korean
13.4K 106100K–1M rows
DatasetShadenA/MathNet

Multilingual math problems with images

Problems from countries and competitions with images and text for QA and solution generation.

  • Question answering
  • Text generation
  • Image → text
  • English
  • Portuguese
  • Spanish
  • +9
13.3K 10510K–100K rows
Datasetwikimedia/wikisource

Cleaned Wikisource texts in many languages

A corpus of cleaned full Wikisource texts in dozens of languages for training and evaluating language models.

  • Text generation
  • Fill mask
  • Arabic
  • Assamese
  • Azerbaijani
  • +9
13.2K 881M–10M rows
Datasetstanfordnlp/SHP

385K human response preferences

385K human preferences for answer helpfulness across 18 topics.

  • Text generation
  • Question answering
  • English
13.1K 333100K–1M rows
Datasetmikex86/stackoverflow-posts

Archive of 60 million StackOverflow posts in Markdown

A large corpus of StackOverflow questions, answers, and code up to June 14, 2023, in Markdown text.

  • Question answering
  • Text generation
  • Code
  • English
13.1K 6410M–100M rows
Datasetallenai/qasper

Questions and answers on NLP research papers

1,585 NLP papers and 5,049 reader questions with practitioner answers for QA training and evaluation.

  • Question answering
  • English
13.1K 1291K–10K rows
DatasetTeam-ACE/ToolACE

Dialogues for training models on API usage

Agent dialogues with tool calls from a pool of 26,507 APIs — for training function calling models.

  • Text generation
  • English
  • Chinese
13K 21210K–100K rows
Datasetgoogle/xquad

Multilingual question-answering benchmark

A parallel corpus of 1,190 question-answer pairs across 10+ languages for cross-lingual QA evaluation.

  • Question answering
  • Arabic
  • German
  • Greek
  • +9
12.9K 4310K–100K rows
DatasetOpenLLM-France/Lucie-Training-Dataset

Multilingual text corpus for pre-training Lucie-7B LLM

Tens of billions of text examples in five European languages and code for pre-training base language models.

  • Text generation
  • English
  • French
  • German
  • +3
12.9K 4010B–100B rows
Datasetopenbmb/UltraData-SFT-Agent-2609Top pick

500k agent trajectories for SFT

English-Chinese tool use trajectories for instruction-tuning agent models, approx. 500k examples.

  • Text generation
  • Question answering
  • English
  • Chinese
12.9K 324100K–1M rows
DatasetJailbreakV-28K/JailBreakV-28k

JailBreakV-28K: MLLM jailbreak benchmark

A collection of images and texts for testing multimodal LLM robustness against jailbreak attacks.

  • Text generation
  • Question answering
12.9K 9210K–100K rows
Datasetgoogle/wmt24ppTop pick

WMT24++: Translation into 55 Languages

Human translations and post-editing for 55 en→xx pairs from the WMT24++ benchmark.

  • Translation
  • Arabic
  • Bulgarian
  • Bangla
  • +9
12.7K 9710K–100K rows
Datasetopencsg/chinese-fineweb-edu

Chinese FineWeb Educational Corpus

Chinese pre-training corpus of 10-100 million texts selected for educational NLP tasks.

  • Text generation
  • Chinese
12.5K 11910M–100M rows
DatasetTIGER-Lab/WebInstruct-verified

Verified instructions for reasoning

A set of English text questions and instructions for training LLM reasoning.

  • Question answering
  • English
12.5K 69100K–1M rows
Datasetgoogle-research-datasets/tydiqa

TyDi QA multilingual question answering dataset

204 thousand question-answer pairs from Google Research in 11 typologically diverse languages.

  • Question answering
  • Arabic
  • Bangla
  • English
  • +8
12.4K 39100K–1M rows
DatasetJoschka/big_bench_hard

BIG-Bench Hard dataset for LLM evaluation

BIG-Bench Hard data adapted for integration into the Inspect AI evaluation framework.

  • Question answering
  • Multiple choice
  • Text generation
  • English
12.4K 41K–10K rows
DatasetCohereLabs/aya_collection

Multilingual Aya collection for NLP tasks

A large multilingual collection of texts for classification, summarization, and translation.

  • Text classification
  • Summarization
  • Translation
  • Acehnese
  • Afrikaans
  • Amharic
  • +9
11.9K 252100M–1B rows