EleutherAI

EleutherAI is a non-profit artificial intelligence research lab focused on interpretability and alignment of large models.

Introduction to EleutherAI organization

Founded in July 2020 by Connor Leahy, Sid Black, and Leo Gao, EleutherAI has grown from a Discord server discussing GPT-3 to a leading non-profit research organization focused on large-scale artificial intelligence research. Our work has historically focused on supporting the use of cutting-edge artificial intelligence technologies by training and publishing models and promoting open scientific norms for natural language processing. As public access to large-scale pre-trained AI models significantly increases, our focus is shifting to research on AI explainability and alignment.

EleutherAI operates primarily through a public Discord server, where research in the field is discussed and our projects are coordinated. EleutherAI adopts an open and collaborative research model, and EleutherAI's Discord server does not strictly distinguish between employees, volunteers, and collaborators from other institutions. However, EleutherAI's community is dedicated to researchers and research-level discussions, and EleutherAI requires primary observation from people interested in learning about artificial intelligence research.

EleutherAI employs a number of full- and part-time researchers who work alongside a dozen regular volunteers and external collaborators.

Eleuther's main results

EleutherAI has released many results related to large models, including various data sets and large models.

The large models released by EleutherAI are listed below:

Model Name Model Introduction DataLearner Model Information Card Address

Pythia series The Pythia series contains 16 major models and 154 training result checkpoints. It is a series of large language models released by EleutherAI in February 2023. Pythia model information card

GPT-JGPT-J is a 6 billion parameter open source English autoregressive language model trained on Pile. At the time of release, it was the world's largest public GPT-3-style language model. GPT-J 6B model information card

The GPT-NeoX series uses the GPT-NeoX library to train large models containing multiple versions, with a maximum parameter scale of 20 billion. GPT-NeoX model information card

Other information about EleutherAI

EleutherAI’s official website: https://www.eleuther.ai/

EleutherAI’s GitHub address: https://github.com/EleutherAI

EleutherAI’s HuggingFace address: https://huggingface.co/EleutherAI