DeepSpeed Chat
lastmod 2023-05-02created 2023-04-181 edits
A framework for training ChatGPT-like models using RLHF
Features
- Fast, affordable, scalable open system framework
- Can produce high-quality ChatGPT-style models of any scale through End-to-End RLHF (Reinforcement Learning Human Feedback)
- With 1 click, can train, generate, and serve a 1.3B-parameter ChatGPT model within 1.36 hours on a single NVIDIA A6000 GPU with 48GB of memory
- Used by Databricks Dolly, CarperAI-TRLX, Huggingface-PEFT, and others