EleutherAI
A non-profit research lab that builds openly licensed AI models and training datasets, run largely by volunteers.
EleutherAI started in 2020 as a group of volunteers on Discord who wanted to replicate the large language models that companies were keeping private, and grew into a registered non-profit research institute. Its early releases, GPT-Neo and GPT-J, were among the first open alternatives to GPT-3, and its Pile dataset became a standard ingredient in open model training.
Its more useful work now is often the unglamorous kind: cleaning up the raw material everyone else builds on. Common Pile, released in 2025, is the largest training corpus assembled entirely from openly licensed text. In 2026 the lab worked with Hugging Face on FineBooks, which tests whether modern text recognition models can produce cleaner versions of the millions of scanned public-domain books that open datasets rely on.