Sunsimiao is a Chinese medical large language model designed to provide safe, reliable, and accessible healthcare information. It targets medical professionals, students, and the general public seeking trustworthy medical knowledge in Chinese. The project aims to leverage extensive medical data to improve diagnostic support, exam preparation, and general medical Q&A.

How It Works

Sunsimiao models are fine-tuned from foundational models like Qwen2-7B using a proprietary, continuously updated medical dataset. This dataset comprises medical literature, textbooks, multi-specialty diagnostic data, patient consultation dialogues, and medical knowledge Q&A, curated through both open-source data and GPT-4 generation, followed by manual cleaning and automated analysis. This approach allows the model to achieve state-of-the-art performance on medical benchmarks.

Quick Start & Requirements

• Install/Run: Uses ModelScope pipelines.

  from modelscope.pipelines import pipeline
  from modelscope.utils.constant import Tasks
  pipe = pipeline(task=Tasks.text_generation, model='AI-ModelScope/Sunsimiao', model_revision='v1.0.0')
  result = pipe("小孩发烧了怎么办？")
  

• Prerequisites: Python, ModelScope library. No specific hardware requirements mentioned for basic inference, but fine-tuning would require significant GPU resources.
• Links: ModelScope Studio, WiseModel

Highlighted Details

• Sunsimiao-7B achieves SOTA on the CMB-Exam benchmark, with 79.23% accuracy on the Chinese National Practicing Physician Qualification Exam.
• Models are fine-tuned from Qwen2-7B and other base models.
• A comprehensive medical dataset is continuously updated.
• Offers multiple model versions including 7B, 0.1M, and 6B parameter sizes.

Maintenance & Community

The project is initiated by Associate Professor Dong Xue from East China Normal University. It receives support from OpenI Cloud, ModelScope, OpenXLab, and Huggingface. Technical guidance and compute resources were provided by the Shanghai Artificial Intelligence Laboratory's Sheng-Pu Yu Large Model Practice Camp and InternStudio.

Licensing & Compatibility

The project's licensing is not explicitly stated in the README. The models are provided "as is." Compatibility for commercial use or closed-source linking is not detailed.

Limitations & Caveats

The models have inherent limitations and may produce incorrect, harmful, or offensive outputs. Users are cautioned against using the models for critical or high-risk decision-making without careful consideration. The project is based on fine-tuning other models, and liability is disclaimed for any issues arising from its use.