Research institute · Research orgs · Washington, DC · Est. 2020
EleutherAI is an open research organization that develops and releases open-weight large language models, datasets, and model evaluation tooling, with an emphasis on mechanistic interpretability and model evaluation. The group’s best-known open artifacts include the GPT-Neo model family, GPT-J, GPT-NeoX-20B, The Pile pretraining dataset, the Pythia training/checkpoint suite for interpretability and training-dynamics work, and the lm-evaluation-harness framework used to run standardized evaluations across language models.
No curated programs yet. Check the careers page ↗.
EleutherAI published GPT-NeoX-20B and released training/evaluation code and model weights in an open-weight form via its repository for public use and evaluation.
Pythia: interpretability across training and scaling (suite + code/data released)EleutherAI released Pythia as a suite for analyzing large language models across training and scaling, with associated code and checkpoints made available for interpretability research.
EleutherAI’s nonprofit status (EIN and 501(c)(3) designation)EleutherAI’s nonprofit entity, EleutherAI Institute (EIN 92-2215190), is listed as a 501(c)(3) with tax exemption issued in Nov 2023.