Baize is a family of open-source chat models ranging from 7B to 30B parameters, trained using LoRA technique for parameter-efficient fine-tuning. The training data consists of 100k dialogs generated by ChatGPT interacting with itself, supplemented with Alpaca dataset. Models are released as LoRA weights that can be merged into standard HuggingFace format for deployment anywhere HF is supported. The repository provides CLI and API interfaces via FastChat integration.