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

Datasetgoogle-research-datasets/paws-x

Multilingual sentence pairs for paraphrase detection

Over 320,000 sentence pairs across six languages for testing and training models in semantic understanding.

  • Text classification
  • German
  • English
  • Spanish
  • +4
7.5K 54100K–1M rows
DatasetLabHC/bias_in_bios

Biography texts for evaluating gender bias

English biographies with profession and gender labels for studying biases in NLP models.

  • Text classification
  • English
7.4K 24100K–1M rows
Datasetjmhessel/newyorker_caption_contest

New Yorker humor understanding benchmark

Cartoons and captions from the New Yorker contest for evaluating humor understanding in models.

  • Image → text
  • Multiple choice
  • Text classification
  • English
7.2K 78100K–1M rows
Datasetfacebook/kilt_tasks

KILT benchmark on Wikipedia

11 datasets for fact-checking, QA, entity linking, and dialogue on a single Wikipedia dump.

  • Fill mask
  • Question answering
  • Text classification
  • English
7.1K 691M–10M rows
Datasetdell-research-harvard/newswire

Historical US wire service news

2.7 million US wire service articles from 1878-1977 with geolocation, topics, and entities.

  • Text classification
  • Text generation
  • Text retrieval
  • English
6.9K 951M–10M rows
Datasetjackhhao/jailbreak-classification

Jailbreak prompt classification

LLM prompts labeled as jailbreak or benign to train harmful request detectors.

  • Text classification
  • English
6.8K 1111K–10K rows
DatasetHello-SimpleAI/HC3

HC3: humans vs. ChatGPT

A corpus of human and ChatGPT answers for comparison and text detection.

  • Text classification
  • Question answering
  • Sentence similarity
  • English
  • Chinese
6.7K 24410K–100K rows
DatasetHuggingFaceH4/orca_dpo_pairs

OpenOrca-based preference pairs for DPO tuning

Preprocessed English dataset for language model alignment using DPO.

  • Text classification
  • Token classification
  • Table question answering
  • English
6.7K 3210K–100K rows
Datasetkakaobrain/coyo-700m

747M Image-Text Pairs

A large corpus of alt-text and image pairs from HTML for training multimodal models.

  • Text → image
  • Image → text
  • Zero shot classification
  • English
6.5K 177100M–1B rows
DatasetLLM-Digital-Twin/Twin-2K-500

Profiles of 2058 US residents for digital twins

Demographic and psychological data from a representative sample of people for character simulation using LLMs.

  • Text classification
  • Text generation
  • Question answering
  • English
6.4K 431K–10K rows
Datasetgretelai/gretel-pii-masking-en-v1

Synthetic documents with PII annotations

Generated texts from various domains with PII and PHI entity annotations for training anonymization models.

  • Text classification
  • Text generation
  • English
6.2K 4710K–100K rows
Datasetklue/klue

Korean NLU benchmark KLUE

8 Korean NLU tasks for evaluating language models: from NER and NLI to MRC and DST.

  • Fill mask
  • Question answering
  • Text classification
  • Korean
6K 99100K–1M rows
Datasetproj-persona/PersonaHubTop pick

PersonaHub: Personas for Synthetic Data

A collection of personas for generating diverse synthetic texts and instructions.

  • Text generation
  • Text classification
  • Token classification
  • English
  • Chinese
5.8K 834100K–1M rows
DatasetBabelscape/wikineural

Multilingual corpus for named entity recognition

From 1 to 10 million annotated texts in six European languages for the NER task, collected using BabelNet and BERT.

  • Token classification
  • German
  • English
  • Spanish
  • +6
5.4K 381M–10M rows
DatasetBeIR/msmarco

English Text Retrieval Benchmark (BEIR)

A subset of MS MARCO in the BEIR benchmark format of millions of English documents and search queries.

  • Zero shot classification
  • Text retrieval
  • English
5.3K 161M–10M rows
DatasetHFUT-MMH-Lab/multimodal_mental_health_dataset_MMH

Multimodal dataset for mental disorder diagnosis

Participant audio, video, and texts in English and Chinese with clinically confirmed diagnoses for distinguishing disorders.

  • Audio classification
  • Video classification
  • Text classification
  • English
  • Chinese
5.3K 510K–100K rows
DatasetTHUIR/Qilin

Multimodal user session dataset for RAG and search

Large Chinese dataset (1M to 10M) for search, recommendations, and RAG systems.

  • Question answering
  • Text classification
  • Sentence similarity
  • Chinese
5.3K 441M–10M rows
Datasetmteb/tweet_sentiment_extraction

Tweet sentiment extraction for MTEB

A set of English tweets from a Kaggle competition for evaluating and training text embeddings in the sentiment analysis task.

  • Text classification
  • English
5.1K 3810K–100K rows
DatasetAnthropic/model-written-evals

Evaluations written by language models

A collection of eval tasks generated by LLMs to test model behavior.

  • Multiple choice
  • Zero shot classification
  • Question answering
  • English
5K 711K–10K rows
Datasetai4privacy/pii-masking-200k

PII masking on 200k examples

Texts with personally identifiable information in English, French, German, and Italian for privacy masking training.

  • Text classification
  • Token classification
  • Table question answering
  • English
  • French
  • German
  • +1
4.7K 132100K–1M rows
DatasetM-A-D/Mixed-Arabic-Datasets-Repo

Large Arabic Text Corpus

A large-scale collection of diverse Arabic texts from various sources for a wide range of NLP tasks.

  • Text classification
  • Question answering
  • Translation
  • Arabic
4.5K 39100M–1B rows
Datasetai4privacy/pii-masking-300k

PII personal data masking

300k texts across 6 languages with PII annotations for training data removal models.

  • Text classification
  • Token classification
  • Table question answering
  • English
  • French
  • German
  • +3
4.2K 121100K–1M rows
Datasetdeclare-lab/HarmfulQA

LLM safety benchmark and dataset

A set of English prompts and responses generated via ChatGPT to evaluate model vulnerabilities and alignment.

  • Text generation
  • Text classification
  • English
4K 491K–10K rows