DoctorGLM is a Chinese medical question-answering model fine-tuned from ChatGLM-6B. It targets medical professionals and researchers seeking a specialized conversational AI for healthcare inquiries, offering improved reliability and multi-turn dialogue capabilities.

How It Works

DoctorGLM leverages ChatGLM-6B as its base, fine-tuning it on a large corpus of Chinese medical dialogues and Q&A pairs across various departments. The project explores both LoRA and P-Tuning v2 methods for fine-tuning, with a recent emphasis on P-Tuning v2 for enhanced multi-turn dialogue and model reliability. Quantized versions (INT4, INT8) are available for reduced memory footprint, though the README notes current performance issues with quantization.

Quick Start & Requirements

• Install: pip install deep_training cpm_kernels icetk transformers>=4.26.1 torch >=1.12.0
• Prerequisites: Python 3.12+, CUDA >= 12 (recommended for GPU), >= 13GB VRAM for unquantized models. Initial run downloads ChatGLM-6B weights.
• Resources: Quantized models require 6-8GB VRAM. Setup involves downloading weights and potentially running Jupyter notebooks for inference or Gradio deployment.
• Links: Project Page, Arxiv, Online Demo

Highlighted Details

• Fine-tuned on 1.9M Chinese medical Q&A pairs across multiple departments.
• Supports multi-turn conversations using P-Tuning v2.
• Offers Gradio-based deployment for easy online interaction and parameter tuning.
• Includes comparison results against base ChatGLM, noting potential "repeating" issues.

Maintenance & Community

The project was last updated on April 18, 2023, with planned updates including referencing sources in dialogue and uploading to Hugging Face. No specific community channels (Discord/Slack) are listed.

Licensing & Compatibility

The project's licensing is not explicitly stated in the README. Compatibility for commercial use or closed-source linking is not specified.

Limitations & Caveats

Quantized models (INT4/INT8) are noted to have significant performance issues. The project acknowledges "repeating" issues in generated responses based on initial tests. The README mentions that LoRA fine-tuning was abandoned due to dialogue capability loss.