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

Datasettiiuae/falcon-refinedweb

English web corpus from CommonCrawl

A filtered and deduplicated English web corpus from CommonCrawl for training language models.

  • Text generation
  • English
25.3K 979100M–1B rows
Datasetdatabricks/officeqa-pro-v2

OfficeQA Pro v2 document reasoning benchmark

Dataset for evaluating models and agents on US Federal report documents.

  • Question answering
  • Text generation
  • Text retrieval
  • English
25.2K 25up to 1K rows🔒 on request
Datasetstanford-star/relbench-v1

Relational databases for RelBench v1 benchmark

A collection of related parquet tables and task descriptions from Stanford for training neural networks on relational databases.

    25.1K 6up to 1K rows
    Datasetgoogle/deepsearchqaTop pick

    Multi-step information search benchmark

    900 prompts for evaluating agents on complex multi-step information search tasks across 17 domains.

    • Question answering
    • English
    25.1K 137up to 1K rows
    Datasetallenai/dolma3_poolTop pick

    Dolma 3 raw text pool with 9 trillion tokens

    A massive unfiltered English corpus of web pages, research papers, and code for pre-training large language models.

    • Text generation
    • English
    25K 4210B–100B rows
    Datasetzgcagi/ZGCM-1-Data

    ZGCM-1 math and search dataset

    Large bilingual text dataset for mathematical reasoning and agentic search.

    • Text generation
    • Question answering
    • Chinese
    • English
    25K 971B–10B rows🔒 on request
    Datasetdi-zhang-fdu/AIME_1983_2024

    AIME Math Competition Benchmark (1983-2024)

    Collection of AIME math competition problems from 1983-2024 for unbiased LLM reasoning evaluation.

      24.9K 41up to 1K rows
      Datasetcardiffnlp/databench

      Tabular QA benchmark

      80 tabular datasets and questions in English and Spanish for LLM evaluation.

      • Table question answering
      • Question answering
      • English
      • Spanish
      24.8K 161K–10K rows
      DatasetSingleBicycle/4KLSDBTop pick

      4KLSDB dataset for restoration and generation

      A native 4K image dataset for restoration and text-to-image generation tasks.

      • Image → image
      • Text → image
      • English
      24.8K 18100K–1M rows
      DatasetCohereLabs/Global-MMLU-Lite

      Global-MMLU-Lite Multilingual Benchmark

      Compact version of the Global-MMLU benchmark by Cohere Labs across 23 languages for fast LLM evaluation.

      • Arabic
      • Bangla
      • Czech
      • +9
      24.6K 4510K–100K rows
      DatasetKodCode/KodCode-V1-SFT-4o

      Synthetic coding tasks and tests for SFT

      Large open-source English synthetic dataset with verifiable solutions and programming tests.

      • Question answering
      • English
      24.5K 10100K–1M rows
      Datasetallenai/dolmino-mix-1124

      OLMo2 Annealing Data Mix

      High-quality mix of web texts, articles, encyclopedias, and code for the second training stage of OLMo2.

      • Text generation
      • English
      24.4K 105100M–1B rows
      DatasetArtificialAnalysis/AA-LCRTop pick

      AA-LCR Long-Context Logic Benchmark

      100 complex questions over large document sets (~100k tokens) to test deep text analysis.

        24.4K 40up to 1K rows
        DatasetMedOtter/msd-liver

        MSD liver segmentation dataset

        CT scans with liver and tumor annotations from the Medical Segmentation Decathlon challenges.

        • Image segmentation
        24.4K 0up to 1K rows
        DatasetProgramGym/programgym-1648

        Program reconstruction tasks from behavior

        A set of 1,648 engineering tasks for training AI agents to reconstruct source code from executables.

        • Text generation
        24.3K 41K–10K rows
        Datasetarsaporta/symile-m3

        Symile-M3 multilingual audio-visual-text dataset

        Multilingual dataset of audio, images, and text for evaluating multimodal models.

        • Zero shot classification
        • Zero shot image classification
        • Arabic
        • Greek
        • English
        • +7
        24.3K 810M–100M rows
        DatasetOpenSourceAP/CrossSection

        CrossSection asset pricing signals

        A large-scale financial dataset with market signals and portfolio returns for quantitative analysis.

        • English
        24.3K 31M–10M rows
        Datasetnvidia/Aegis-AI-Content-Safety-Dataset-2.0

        LLM response safety: 33k dialogues

        33,416 annotated human-LLM interaction examples for classifying unsafe content.

        • Text classification
        • English
        24.3K 12010K–100K rows
        Datasettiyunes/svg-primitive-labels-sam3-v1-filtered

        Filtered SVG Primitive Labels (SAM3 v1)

        Over 1.16M cleaned query-SVG markup pairs with source XML, rasters, and validity masks.

        • Other
        • English
        24.3K 01M–10M rows🔒 on request
        DatasetIntelligenceLab/Long-Horizon-Terminal-Bench

        Long-Horizon Terminal Tasks Benchmark

        LHTB evaluates how LLM agents maintain useful work in a containerized terminal over hundreds of steps.

        • Text generation
        • English
        24.3K 144up to 1K rows