CMLM-ZhongJing is a large language model specifically designed for Traditional Chinese Medicine (TCM). It aims to leverage ancient wisdom with modern AI to create a professional tool for the medical field, offering insights into TCM knowledge. The project provides fine-tuned weights for existing models, enabling TCM practitioners and researchers to explore AI applications in their domain.

How It Works

The project fine-tunes existing LLMs (Baichuan2-13B-Chat, Qwen1.5-1.8B-Chat) on a proprietary TCM dataset. A key innovation is their "multi-task treatment behavior decomposition instruction construction strategy," which generates diverse and specialized instruction data by breaking down TCM diagnostic and treatment processes into 15 distinct scenarios. This approach, inspired by human memory and learning, aims to improve the model's understanding and reasoning capabilities in the low-error-tolerance domain of TCM, mitigating the risk of factual inaccuracies common in self-instruct methods.

Quick Start & Requirements

• Install/Run: Fine-tuned weights are available on Huggingface. A Gradio-based WebDemo can be launched locally via python WebDemo.py.
• Prerequisites: The 1.8B model can run on a single Tesla T4 GPU. Specific hardware requirements for larger models are not detailed.
• Resources: The project utilizes a large dataset (26GB+ of tokens).
• Links: Huggingface, WebDemo

Highlighted Details

• Offers fine-tuned weights for 13B (Baichuan2-13B-Chat) and 1.8B (Qwen1.5-1.8B-Chat) models.
• Employs a novel instruction construction strategy using specialized prompts and TCM data tables to generate high-quality, domain-specific training data.
• Demonstrates promising generalization capabilities across various TCM clinical specialties beyond gynecology in preliminary tests.
• Includes a comparative analysis against other LLMs (Wenxin Yiyan, Spark) on TCM diagnostic and prescription tasks.

Maintenance & Community

The project is guided by professors from Fuzhou University of Technology, Health Yangtze River Delta Research Institute, and Fudan University. It acknowledges contributions from various physicians and researchers for data support, annotation, and evaluation. Collaboration is welcomed from TCM practitioners.

Licensing & Compatibility

The project states: "This research is for academic research use only and may not be used for commercial purposes without permission. It may not be used for clinical practice in medical scenarios or scenarios with potential medical intent." This implies a non-commercial license, though the specific license type is not explicitly stated in the README.

Limitations & Caveats

The model is in the laboratory testing phase and is not yet suitable for clinical practice or commercial use. Its diagnostic and prescription generation capabilities are considered preliminary and not highly reliable for clinical decision-making. The project explicitly states it does not possess medical practice capabilities.