ColossalChat
lastmod 2023-05-02created 2023-04-031 edits
An open-source RLHF pipeline solution for cloning ChatGPT
Features
- Based on the LLaMA model
- Supervised data collection
- Supervised fine-tuning
- Reward model training
- Reinforcement Learning fine-tuning
- Interactive demo that runs online
- Open-source complete RLHF training code, including 7B/13B models
- 104k bilingual dataset in Chinese/English
- 4-bit quantization of the 7B model. Requires only 4GB of GPU memory
- Model weights included. Easily reproducible on a single server
- Large models/datasets/optimizations and more will continue to be added