EleutherAI

EleutherAI is a nonprofit AI research lab that creates open training data (The Pile, Common Pile), the GPT-NeoX training framework, the LM Evaluation Harness evaluation tool, and the Pythia model suite, which is released even with its checkpoints. It studies how a model's internals develop during training, and develops safety techniques that persist even in open-weight models (such as Deep Ignorance). It conducts research on a public Discord server and runs programs such as the SOAR mentoring program and the EvalEval Coalition.

• Public training datasets The Pile and Common Pile
• Training framework GPT-NeoX
• LM Evaluation Harness, a standard tool for evaluating open models
• Pythia model suite, which releases intermediate training checkpoints and data order
• Deep Ignorance, which builds safeguards that resist fine-tuning through pretraining data filtering
• Bergson, an open-source library for data attribution analysis

Not applicable (research organization/company)

No usage fee for the lab itself; billed separately through its commercial products and API

Estimate based on AI knowledge - please verify on the official site

• One of 56 AI tools in the Frontier Labs field.
• Connected to 5 other tools — #16 in its field.