MedGo: Medical Large Language Model Based on Qwen3-32B

English | 简体中文

📋 Table of Contents

• Introduction
• Key Features
• Performance
• Quick Start
• Training Details
• Use Cases
• Limitations & Risks
• Citation
• License
• Contributing
• Contact

🎯 Introduction

MedGo is a general-purpose medical large language model fine-tuned from Qwen3-32B, designed for clinical medicine and research scenarios. The model is trained on large-scale multi-source medical corpora and enhanced with complex case data, supporting various capabilities including medical Q&A, clinical summary, clinical reasoning, multi-turn dialogue, and scientific text generation.

🌟 Core Capabilities

• 📚 Medical Knowledge Q&A: Professional responses based on authoritative medical literature and clinical guidelines
• 📝 Clinical Documentation: Automated medical record summaries, diagnostic reports, and medical documentation
• 🔍 Clinical Reasoning: Differential diagnosis, examination recommendations, and treatment suggestions
• 💬 Multi-turn Dialogue: Patient-doctor interaction simulation and complex case discussions
• 🔬 Research Support: Literature summarization, research idea generation, and quality control review

✨ Key Features

📊 Performance

MedGo demonstrates excellent performance across multiple medical and general evaluation benchmarks, showing competitive results among 32B-parameter models:

Key Benchmark Results

• AIMedQA: Medical question answering comprehension
• CME: Clinical reasoning evaluation
• DiagnosisArena: Diagnostic capability assessment
• MedQA / MedMCQA: Medical multiple-choice questions
• PubMedQA: Biomedical literature Q&A
• MMLU-Pro: Comprehensive capability evaluation

Performance Highlights:

• ✅ Average Score: ~70 points (excellent performance in the 32B parameter class)
• ✅ Strong Tasks: Clinical reasoning (DiagnosisArena, CME) and multi-turn medical Q&A
• ✅ Balanced Capability: Good performance in medical semantic understanding and multi-task generalization

🚀 Quick Start

Requirements

• Python >= 3.8
• PyTorch >= 2.0
• Transformers >= 4.35.0
• CUDA >= 11.8 (for GPU inference)

Installation

# Clone the repository
git clone https://github.com/OpenMedZoo/MedGo.git
cd MedGo

# Install dependencies
pip install -r requirements.txt

Model Download

Download model weights from HuggingFace:

# Using huggingface-cli
huggingface-cli download OpenMedZoo/MedGo --local-dir ./models/MedGo

# Or using git-lfs
git lfs install
git clone https://huggingface.co/OpenMedZoo/MedGo

Basic Inference

from transformers import AutoModelForCausalLM, AutoTokenizer

# Load model and tokenizer
model_path = "OpenMedZoo/MedGo"
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_path,
    device_map="auto",
    trust_remote_code=True,
    torch_dtype="auto"
)

# Medical Q&A example
messages = [
    {"role": "system", "content": "You are a professional medical assistant. Please answer questions based on medical knowledge."},
    {"role": "user", "content": "What is hypertension and what are the common treatment methods?"}
]

# Generate response
inputs = tokenizer.apply_chat_template(
    messages, 
    tokenize=True, 
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)

outputs = model.generate(
    inputs,
    max_new_tokens=512,
    temperature=0.7,
    top_p=0.9,
    do_sample=True
)

response = tokenizer.decode(outputs[0][len(inputs[0]):], skip_special_tokens=True)
print(response)

Batch Inference

# Use the provided inference script
python scripts/inference.py \
    --model_path OpenMedZoo/MedGo \
    --input_file examples/medical_qa.jsonl \
    --output_file results/predictions.jsonl \
    --batch_size 4

Accelerated Inference with vLLM

from vllm import LLM, SamplingParams

# Initialize vLLM
llm = LLM(model="OpenMedZoo/MedGo", trust_remote_code=True)
sampling_params = SamplingParams(temperature=0.7, top_p=0.9, max_tokens=512)

# Batch inference
prompts = [
    "What are the symptoms and treatment methods for diabetes?",
    "What dietary precautions should hypertensive patients take?"
]

outputs = llm.generate(prompts, sampling_params)
for output in outputs:
    print(output.outputs[0].text)

🔧 Training Details

MedGo employs a two-stage fine-tuning strategy to balance general medical knowledge with clinical task adaptation.

Stage I: General Medical Alignment

Objective: Establish a solid foundation of medical knowledge and improve Q&A standardization

• Data Sources:

  • Authoritative medical literature (PubMed, medical textbooks)
  • Clinical guidelines and diagnostic standards
  • Medical encyclopedia entries and terminology databases

• Training Methods:

  • Supervised Fine-Tuning (SFT)
  • Chain-of-Thought (CoT) guided samples
  • Medical terminology alignment and safety constraints

Stage II: Clinical Task Enhancement

Objective: Enhance complex case reasoning and multi-task processing capabilities

• Data Sources:

  • Real medical records (fully anonymized)
  • Outpatient and emergency records with complex multi-diagnosis samples
  • Research articles and quality control cases

• Data Augmentation Techniques:

  • Semantic paraphrasing and multi-perspective expansion
  • Complex case synthesis
  • Doctor-patient interaction simulation

• Training Methods:

  • Multi-Task Learning (medical record summary, differential diagnosis, examination suggestions, etc.)
  • Preference Alignment (DPO/KTO)
  • Expert feedback iterative optimization

Training Optimization Focus

• ✅ Strengthen information extraction and cross-evidence reasoning for complex cases
• ✅ Improve medical consistency and interpretability of outputs
• ✅ Optimize expression compliance and safety
• ✅ Continuous iteration through expert samples and automated evaluation

💡 Use Cases

✅ Suitable Scenarios

🚫 Unsuitable Scenarios

• ❌ Cannot Replace Doctors: Only an auxiliary tool, not a standalone diagnostic basis
• ❌ High-Risk Operations: Not recommended for surgical decisions or other high-risk medical operations
• ❌ Rare Disease Limitations: May perform poorly on rare diseases outside training data
• ❌ Emergency Care: Not suitable for scenarios requiring immediate decisions

⚠️ Limitations & Risks

Model Limitations

1. Understanding Bias: Despite covering extensive medical knowledge, may still produce understanding biases or incorrect recommendations
2. Complex Cases: Higher risk for cases with complex conditions, severe complications, or missing information
3. Knowledge Currency: Medical knowledge continuously updates; training data may lag
4. Language Limitation: Primarily designed for Chinese medical scenarios; performance in other languages may vary

Usage Recommendations

• ⚠️ Use in controlled environments with clinical expert review of generated results
• ⚠️ Treat model outputs as auxiliary references, not final diagnostic conclusions
• ⚠️ For sensitive cases or high-risk scenarios, expert consultation is mandatory
• ⚠️ Deployment requires internal validation, security review, and clinical testing

Data Privacy & Compliance

• 🔒 Training data fully anonymized
• 🔒 Attention to patient privacy protection during use
• 🔒 Production deployment must comply with healthcare data security regulations (e.g., HIPAA, GDPR)
• 🔒 Local deployment recommended to avoid sensitive data transmission

📚 Citation

If MedGo is helpful for your research or project, please cite our work:

@misc{openmedzoo_2025,
	author       = { OpenMedZoo },
	title        = { MedGo (Revision 640a2e2) },
	year         = 2025,
	url          = { https://huggingface.co/OpenMedZoo/MedGo },
	doi          = { 10.57967/hf/7024 },
	publisher    = { Hugging Face }
}

📄 License

This project is licensed under the Apache License 2.0.

Commercial Use Notice:

• ✅ Commercial use and modification allowed
• ✅ Original license and copyright notice must be retained
• ✅ Contact us for technical support when integrating into healthcare systems

🤝 Contributing

We welcome community contributions! Here's how to participate:

Contribution Types

• 🐛 Submit bug reports
• 💡 Propose new features
• 📝 Improve documentation
• 🔧 Submit code fixes or optimizations
• 📊 Share evaluation results and use cases

🙏 Acknowledgments

Thanks to all contributors to the MedGo project:

• Model development and fine-tuning algorithm team
• Data annotation and quality control team
• Clinical expert guidance and review team
• Open-source community support and feedback

Special thanks to:

• Qwen Team for providing excellent foundation models
• All healthcare institutions that provided data and feedback

📧 Contact

• HuggingFace: Model Homepage

Copyright

• Publisher: Tongji University Affiliated East Hospital — Sole Corresponding Author
• Co-developer / Technical Support: Shanghai Shuole Technology Co., Ltd.
• Contact: dongfyy@pudong.gov.cn
• Version: v1.0
• Attribution (required): “Powered by Med-Go 32B, released by Tongji University Affiliated East Hospital (v1.0).”