Datasets

15 trillion tokens of filtered Common Crawl from Hugging Face, with the filtering recipe published.

An open reproduction of the LLaMA training data, 1.2 trillion tokens.

5.85 billion image-text pairs. Trained Stable Diffusion. Taken down in 2023 over illegal content, re-released cleaned in 2024.

825GB from 22 sources, assembled by EleutherAI. Trained GPT-J, GPT-NeoX and Pythia.

750GB of filtered Common Crawl text, released with T5.

100,000 reading comprehension questions on Wikipedia. The reading benchmark BERT saturated.

14 million labelled images in 22,000 categories. The dataset that made deep learning for vision possible.

A free monthly crawl of the web since 2008. The raw material of nearly every language model.

70,000 handwritten digits. The first benchmark every student runs.