hari-q3-8b

🧠 Korean Medical LLM (QA-Finetuned) by Healthcare AI Research Institute of Seoul National University Hospital

Welcome to the official repository of the Korean Medical Large Language Model (LLM) developed by the Healthcare AI Research Institute (HARI) at Seoul National University Hospital (SNUH).

---

🚀 Model Overview

• Model Name: snuh/hari-q3-8b
• Architecture: Large Language Model (LLM)
• Fine-tuning Objective: Medical QA (Question–Answer) style generation
• Primary Language: English, Korean
• Domain: Clinical Medicine
• Performance: Achieves 76.78% accuracy on the Korean Medical Licensing Examination (KMLE)
• Key Applications:
  • Clinical decision support (QA-style)
  • Medical education and self-assessment tools
  • Automated medical reasoning and documentation aid

---

📊 Training Data & Benchmark

This model was fine-tuned using a curated corpus of Korean medical QA-style data derived from publicly available, de-identified sources. The training data includes clinical guidelines, academic publications, exam-style questions, and synthetic prompts reflecting real-world clinical reasoning.

• Training Data Characteristics:

  • Focused on Korean-language question–answering formats relevant to clinical settings.
  • Includes guideline-derived questions, de-identified case descriptions, and physician-crafted synthetic queries.
  • Designed to reflect realistic diagnostic, therapeutic, and decision-making scenarios.

• Benchmark Evaluation:

  • KorMedMCQA(5-shot) |Accuracy(%)|gpt-oss-20b|medgemma-27b|hari-q3-8b| |------|---|---|---| |Doctor|0.7333|0.7632|0.7678| |Nurse|0.7722|0.8257|0.8360| |Pharm|0.7684|0.8056|0.8441| |Dentist|0.4316|0.6141|0.6165|

  • MedQA-USMLE(0-shot) ||gpt-oss-20b|medgemma-27b|hari-q3-8b| |------|---|---|---| |Accuracy(%)|0.7777|0.8429|0.8154|

  • JAMA challenge(5-shot) ||gpt-oss-20b|medgemma-27b|hari-q3-8b| |------|---|---|---| |Accuracy(%)|0.7777|0.8429|0.8154|

  • NEJM(5-shot) |Accuracy(%)|gpt-oss-20b|medgemma-27b|hari-q3-8b| |------|---|---|---| |General surgery|0.6809|0.6312|0.6099| |Internal medicine|0.7063|0.6984|0.6587| |Psychiatry|0.7467|0.6933|0.7533| |Pediatrics|0.7374|0.7172|0.6869| |Obgyn|0.5324|0.5683|0.5036|

  • All evaluations were conducted on de-identified, non-clinical test sets, with no real patient data involved.

⚠️ These benchmarks are provided for research purposes only and do not imply clinical safety or efficacy.

---

🔐 Privacy & Ethical Compliance

We strictly adhere to ethical AI development and privacy protection:

• ✅ The model was trained exclusively on publicly available and de-identified data.
• 🔒 It does not include any real patient data or personally identifiable information (PII).
• ⚖️ Designed for safe, responsible, and research-oriented use in healthcare AI.

⚠️ This model is intended for research and educational purposes only and should not be used to make clinical decisions.

---

🏥 About HARI – Healthcare AI Research Institute

The Healthcare AI Research Institute (HARI) is a pioneering research group within Seoul National University Hospital, driving innovation in medical AI.

🌍 Vision & Mission

• Vision: Shaping a sustainable and healthy future through pioneering AI research.
• Mission:
  • Develop clinically useful, trustworthy AI technologies.
  • Foster cross-disciplinary collaboration in medicine and AI.
  • Lead global healthcare AI commercialization and policy frameworks.
  • Educate the next generation of AI-powered medical professionals.

---

🧪 Research Platforms & Infrastructure

• Platforms: SUPREME, SNUHUB, DeView, VitalDB, NSTRI Global Data Platform
• Computing: NVIDIA H100 / A100 GPUs, Quantum AI Infrastructure
• Projects:
  • Clinical note summarization
  • AI-powered diagnostics
  • EHR automation
  • Real-time monitoring via AI pipelines

---

🎓 AI Education Programs

• Basic AI for Healthcare: Designed for clinicians and students
• Advanced AI Research: Targeting senior researchers and specialists in clinical AI validation and deep learning

---

🤝 Collaborate with Us

We welcome collaboration with:

• AI research institutions and medical universities
• Healthcare startups and technology partners
• Policymakers shaping AI regulation in medicine

📧 Contact: [email protected]
🌐 Website: Seoul National University Hospital

---

🤗 Model Usage Example

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

# Load tokenizer and model
model_name = "snuh/hari-q3-8b"

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

prompt = '''
### Instruction:
당신은 임상 지식을 갖춘 유능하고 신뢰할 수 있는 한국어 기반 의료 어시스턴트입니다.
사용자의 질문에 대해 정확하고 신중한 임상 추론을 바탕으로 진단 가능성을 제시해 주세요.
반드시 환자의 연령, 증상, 검사 결과, 통증 부위 등 모든 단서를 종합적으로 고려하여 추론 과정과 진단명을 제시해야 합니다.
의학적으로 정확한 용어를 사용하되, 필요하다면 일반인이 이해하기 쉬운 용어도 병행해 설명해 주세요.

### Question:
60세 남성이 복통과 발열을 호소하며 내원하였습니다.
혈액 검사 결과 백혈구 수치가 상승했고, 우측 하복부 압통이 확인되었습니다.
가장 가능성이 높은 진단명은 무엇인가요?
'''.strip()

messages = [
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=4096
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(response)

---

📄 License

Apache 2.0 License – Free for research and commercial use with attribution.

---

📢 Citation

If you use this model in your work, please cite:

@misc{hari-q3,
    title  = {hari-q3-8b},
    url    = {https://huggingface.co/snuh/hari-q3-8b},
    author = {Healthcare AI Research Institute(HARI) of Seoul National University Hospital(SNUH)},
    month  = {November},
    year   = {2025}
}

---

🚀 Together, we are shaping the future of AI-driven healthcare.

---

Acknowlegments

This work was supported by Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (RS-2025-02653113, High-Performance Research AI Computing Infrastructure Support at the 2 PFLOPS Scale)