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

DatasetKodCode/KodCode-V1-SFT-R1

Synthetic coding tasks with tests

Large corpus of English synthetic coding tasks with verifiable solutions and tests for SFT and RL.

  • Question answering
  • English
61.1K 40100K–1M rows
Datasetnebius/SWE-rebench-V2-PRs

126k real GitHub pull requests in 18+ languages

Large-scale corpus of real GitHub pull requests by Nebius for training and evaluating coding agents.

  • Text generation
  • English
61K 19100K–1M rows
Datasetzai-org/LongBench-v2

LongBench v2 long-context benchmark

An English test set to evaluate how LLMs handle complex long-context tasks.

  • Multiple choice
  • Question answering
  • Text classification
  • English
61K 65up to 1K rows
DatasetOpenSQZ/AutoMathText-V2Top pick

2.46T STEM Tokens for Pre-training

A massive cleaned corpus on STEM, mathematics, code, and reasoning in English and Chinese.

  • Text generation
  • Question answering
  • English
  • Chinese
60.9K 781B–10B rows
DatasetIFM/MegaMathTop pick

Math corpus for LLM pre-training

An open English mathematical dataset exceeding 300 billion tokens for pre-training language models.

  • Text generation
  • English
60.8K 144100M–1B rows
Datasetamolharsh/Ground3D_Dataset

3D scene question-answering with point grounding and metrics

Large English corpus of 3D scenes and dialogues based on ScanNet for training models in spatial and object size understanding.

  • Visual question answering
  • Question answering
  • English
60.5K 11M–10M rows
Datasetnvidia/OpenCodeInstructTop pick

5M instructions for coding LLMs

5 million examples for SFT of coding LLMs: diverse English instructions from NVIDIA.

  • Text generation
  • English
59.1K 1221M–10M rows
DatasetUnipatAI/RoadmapBench

Coding agent benchmark on repository upgrades

115 complex software development tasks from real open-source project updates across five programming languages.

  • Text generation
  • English
56.9K 12up to 1K rows
Datasetharborframework/harbor-mix

A set of 100 complex tasks for evaluating AI agents

A curated selection of 100 complex and diverse agent tasks for fast and inexpensive model evaluation.

  • Text generation
  • Reinforcement learning
56.6K 4up to 1K rows🔒 on request
Datasetbshada/open-schematics

Electronic Schematics and PCB Layouts

Largest collection of schematics and PCBs: sources, images, metadata, and related board files.

  • Text generation
  • Image → text
  • Text → image
  • English
55.7K 20810K–100K rows
Datasetopenlifescienceai/medmcqa

Medical Examination Multiple-Choice Questions

194k medical examination questions with multiple-choice answers in English.

  • Question answering
  • Multiple choice
  • English
55.3K 252100K–1M rows
DatasetHuggingFaceH4/no_robots

10k Human Instructions for SFT

10k instructions and human responses for SFT to help models follow commands better.

  • Text generation
  • English
54.9K 60910K–100K rows
DatasetSkylion007/openwebtext

Open WebText replica for GPT-2 training

Open English replica of OpenAI's WebText used to train GPT-2.

  • Text generation
  • Fill mask
  • English
54.8K 5521M–10M rows
DatasetEthnic-Voices-Tz/ethnic-voices-tanzaniaTop pick

Speech and Texts in Six Tanzanian Languages

A corpus of 462 hours of audio recordings and texts in rare ethnic languages of Tanzania for training ASR and language models.

  • Automatic speech recognition
  • Text generation
54.4K 010K–100K rows
Datasetibm-granite/ChartNet

Multimodal Chart Dataset

A large-scale multimodal chart corpus with images and text for data extraction, VQA, and multilingual understanding.

  • Image → text
  • Visual question answering
  • Table question answering
  • English
  • Arabic
  • Chinese
  • +9
53.6K 6810M–100M rows
Datasetfacebook/belebele

Belebele Multilingual Benchmark

MRC benchmark with questions and 4-way multiple-choice options across 122 language varieties, including low-resource ones.

  • Question answering
  • Zero shot classification
  • Text classification
  • Afrikaans
  • Amharic
  • Arabic
  • +9
53.5K 146100K–1M rows
Datasetopen-thoughts/OpenThoughts3-1.2MTop pick

OpenThoughts3: Reasoning Data

1.2M rows of open data for training reasoning models; third version of OpenThoughts.

  • Text generation
53.5K 2851M–10M rows
Datasetdavidheineman/tiny-rl-refined-tasks-v5

Dataset for SFT and RL training

Text tasks for SFT and RL with prediction tokens and model rewards.

  • Text generation
  • English
53.2K 01M–10M rows
Datasetlhpku20010120/Omni-Edu

Multimodal School Education Instructions

Nearly 70,000 English and Chinese examples for training educational LLMs and VLMs.

  • Text generation
  • Question answering
  • Image text → text
  • English
  • Chinese
52.6K 610K–100K rows
DatasetNEAR-AI/officeqa

US Treasury Bulletin QA (OfficeQA Manifest)

A set of 246 English questions for evaluating QA search systems on financial documents.

  • Question answering
  • English
51.6K 0up to 1K rows
Datasetise-uiuc/Magicoder-OSS-Instruct-75K

OSS-Instruct Code Generation Instructions

A set of OSS-Instruct instructions generated by gpt-3.5-turbo-1106 for code generation tasks.

  • Text generation
51.4K 18610K–100K rows
Datasetthemohal/saraiki-llm-bench

Saraiki LLM evaluation benchmark

A set of seven test tasks to check the quality and safety of language models in the low-resource Saraiki language.

  • Question answering
  • Multiple choice
  • Text generation
  • skr
51K 01K–10K rows🔒 on request
Datasetepfml/FineWeb2-HQ

FineWeb2-HQ: High-quality corpus for pretraining

A quality-filtered subset of FineWeb2 across 20 languages for faster language model pretraining.

  • Text generation
  • Russian
  • Chinese
  • German
  • +9
49.3K 89100M–1B rows
DatasetIFM/Code-Reasoning

Data for reasoning code generation

Coding tasks with reasoning traces, direct answers, and task synthesis in parquet.

  • Text generation
48.9K 106100M–1B rows