Huatuo-Llama-Med-Chinese

Why we included this project

Teams pushing a general-purpose language model into a narrow, high-stakes domain will get a lot from this repo. It walks through the full recipe: building a Chinese medical instruction dataset from a knowledge graph and medical literature, using a stronger model to generate question-answer pairs, then instruction-tuning several open base models with LoRA to keep the training affordable. Along with the code, the authors publish LoRA adapter weights for LLaMA, Alpaca-Chinese, Bloom, and Huozi, so you can load a tuned medical model without reproducing the training run yourself. Inference scripts and prompt templates make it easy to run the variants side by side. If you want a concrete, reproducible answer to how much domain fine-tuning helps a stock model on Chinese medical Q&A, this is a solid place to start.