Baize is an open-source chat model family best known for a clever trick: it taught itself by letting ChatGPT hold conversations with itself. The project fine-tunes Meta’s LLaMA models with LoRA on roughly 100,000 self-generated dialogues, producing capable chatbots that can be trained in hours on a single GPU.
Baize lives on GitHub, where it has collected more than 3,100 stars. Released in 2023 and documented in an arXiv paper, it was one of the influential early open chatbot projects of the LLaMA era. The team generated multi-turn dialogues by seeding ChatGPT with questions from Quora, Stack Overflow, and MedQuAD, letting the model play both sides of the conversation, then combined that self-chat data with Alpaca data for fine-tuning. Trained models were released in 7B, 13B, and 30B sizes, with v2 models published on Hugging Face. The repository contains the full pipeline: data collection scripts, the dialogue datasets, LoRA training code, a script for merging LoRA weights back into standard Hugging Face models, and a demo application. FastChat also added support, giving Baize a CLI and API. The code is GPL-3.0, while the model weights and data are restricted to research use only, with commercial use strictly prohibited. The project is no longer actively developed, but it remains a compact, readable reference for self-instruct fine-tuning.
Baize is free: the code is open source under GPL-3.0 and the weights are downloadable from Hugging Face. Note the license terms: models and data are for research use only, and commercial use is prohibited.
Baize suits LLM researchers, students learning how instruction tuning works, and hobbyists who want a self-contained, historically important example of building a chatbot from self-generated data rather than human annotation.