This repository provides a Chinese medical instruction-tuned version of the ChatGLM-6B model, aimed at improving its performance in the healthcare domain. Researchers and developers working with Chinese medical data can leverage this fine-tuned model for enhanced medical question answering capabilities.

How It Works

The project fine-tunes the ChatGLM-6B model using a custom-built Chinese medical instruction dataset. This dataset is constructed by leveraging a medical knowledge graph (cMeKG) and the GPT-3.5 API to generate diverse question-answer pairs covering diseases, drugs, and examination indicators. The fine-tuning process aims to adapt the base ChatGLM model to understand and respond to medical queries more effectively.

Quick Start & Requirements

• Install dependencies: pip install -r requirements.txt
• Recommended Python version: 3.9+
• Download pre-trained model parameters: Baidu Netdisk or GoogleDrive (links are placeholders, actual links are in the README).
• Interactive testing: python infer.py
• Training requires a GPU with at least 32GB VRAM (A100-SXM-80GB used for training).

Highlighted Details

• Fine-tuned ChatGLM-6B model specifically for Chinese medical domain.
• Utilizes a medical knowledge graph (cMeKG) and GPT-3.5 for dataset creation.
• Demonstrates improved performance over the base ChatGLM-6B on medical Q&A tasks.
• Also trained a medical version of LLaMA, named "Huatuo".

Maintenance & Community

• Developed by Harbin Institute of Technology (HIT) Health Intelligence Group.
• References ChatGLM, ChatGLM-Instruct-Tuning, and CMeKG projects.

Licensing & Compatibility

• Resources are for academic research only.
• Strictly prohibited for commercial use.
• Adherence to third-party code licenses is required.

Limitations & Caveats

The project states that the instruction-tuning method might negatively impact ChatGLM's base capabilities due to the lack of open-source training strategies. Future iterations will focus on open-source models due to copyright considerations. The dataset quality is noted as limited and subject to ongoing iteration. Model-generated content should not be used for actual medical diagnosis.