CMLM-ZhongJing

Why we included this project

Fine-tuned on top of Baichuan2-13B and Qwen1.5-1.8B, this model family is one of the few open efforts aimed squarely at Traditional Chinese Medicine. Its training data amounts to more than 135,000 instruction samples across 15 clinical scenarios, drawing on classical texts and herbal formulas along with the syndrome and tongue pulse patterns used in diagnosis, and five practicing TCM physicians graded the outputs on several quality dimensions. Both a 13B checkpoint and a lightweight 1.8B version are available; the smaller one runs on a single T4 GPU, and a Gradio demo plus a Colab notebook let you try multi-turn conversation without setting anything up locally. Be clear on scope: this is an academic research release whose results are explicitly not medical advice, so treat it as a reference for domain adaptation or research prototyping, not production decision support.