PULSE is a suite of Chinese medical large language models designed for a wide range of healthcare NLP tasks, including health education, exam preparation, report interpretation, and clinical note structuring. It offers pre-trained models and fine-tuning capabilities, targeting researchers and developers in the medical domain.

How It Works

PULSE models are fine-tuned on a substantial dataset of Chinese medical and general domain instructions. The project leverages existing large language models like BLOOMZ-7b1-mt and InternLM-20B as base models, enhancing them with medical-specific knowledge and task-oriented capabilities. This approach allows for rapid development and deployment of specialized medical AI services.

Quick Start & Requirements

• Install: Clone the repository and create a conda environment using llm.yml.
• Prerequisites: Python, PyTorch, Transformers (versions not lower than recommended).
• Hardware:
  • PULSE-7B: 14GB VRAM (FP16), 6GB VRAM (INT4)
  • PULSE-20B: 40GB VRAM (FP16), 12GB VRAM (INT4)
• Demo: Run web_demo_gradio.py for a web UI or cli_demo.py for a command-line interface.
• More Info: PULSE-EVAL, OpenCompass MedBench

Highlighted Details

• Evaluated using Elo rating against models like GPT-4 and ChatGPT on various medical benchmarks (MedQA, PromptCBLUE, etc.).
• Offers fine-tuning scripts via the LLaMA-Factory project (PULSE-tuner).
• Provides quantization solutions for both PULSE-7b (GPTQ) and PULSE-20b (LMDeploy).
• Supports multimodal applications through integration with X-ray models (XrayPULSE).

Maintenance & Community

• Developed by openmedlab, with contributions from Shanghai AI Laboratory, Shanghai Jiao Tong University, and East China University of Science and Technology.
• Related projects include XrayPULSE, PULSE-COVID-19, and tools for medical record structuring and terminology normalization.

Licensing & Compatibility

• Code: Apache 2.0
• Model Weights: GNU AGPL 3.0
• The AGPL 3.0 license for model weights may impose restrictions on commercial use or linking with closed-source applications.

Limitations & Caveats

The models are intended for research purposes only and should not replace professional medical advice for diagnosis or treatment. The accuracy and completeness of information provided cannot be guaranteed.