DISC-MedLLM
| Property | Value |
|---|---|
| Base Model | Baichuan-13B-Base |
| Language | Chinese |
| License | Apache 2.0 |
| Paper | Research Paper |
What is DISC-MedLLM?
DISC-MedLLM is a specialized medical domain large language model developed by Fudan-DISC lab, specifically designed for conversational healthcare scenarios. Built upon the Baichuan-13B architecture, it has been extensively trained on over 470,000 medical dialogue examples and knowledge graph data to provide reliable medical consultations and treatment information.
Implementation Details
The model leverages a comprehensive training dataset (DISC-Med-SFT) that combines real-world medical conversations, knowledge graph-derived QA pairs, and carefully curated behavioral preference data. It implements a goal-oriented strategy that integrates both LLM capabilities and human-in-the-loop feedback mechanisms.
- Training data includes 420k real-world conversations, 50k knowledge graph QA pairs, and 2k human-preference aligned examples
- Utilizes PyTorch framework with text-generation-inference capabilities
- Supports both full parameter and quantized inference options
Core Capabilities
- Knowledge-intensive medical consultations with high reliability
- Multi-turn inquiry support for detailed medical discussions
- Human preference alignment for natural conversation flow
- Integration with medical knowledge graphs for accurate information
Frequently Asked Questions
Q: What makes this model unique?
DISC-MedLLM stands out for its specialized focus on medical conversations and its comprehensive training approach that combines real-world medical dialogues with structured knowledge graph data. The model's architecture ensures both accuracy and natural conversation flow in healthcare scenarios.
Q: What are the recommended use cases?
The model is specifically designed for medical consultations, treatment inquiries, and health support services. However, it should be used for research and testing purposes only, not as a replacement for professional medical advice.
