snuh/specialist-level_medical_knowledge_dataset_sft

specialist-levelmedicalknowledgedatasetsft

Dataset Description

• —Domain: 13 Medical Specialties (Surgery, Psychiatry, Neurology, Dermatology, Medical Law, etc.)
• —Language: Korean, English
• —Task: Clinical Reasoning, Question Answering
• —Augmentation Method: Evol-Instruct (Add Constraints only)
• —Source Data: AI-Hub Specialized Medical Knowledge Data (전문 의학지식 데이터)

Dataset Summary

specialist-level_medical_knowledge_dataset_sft is an integrated collection of augmented SFT data across 13 distinct medical domains, developed by the Healthcare AI Research Institute (HARI) at SNUH.

This dataset is derived and augmented from the [Specialized Medical Knowledge Data (전문 의학지식 데이터)](https://aihub.or.kr/aihubdata/data/view.do?currMenu=115&topMenu=100&&srchDataRealmCode=REALM006&aihubDataSe=data&dataSetSn=71874) provided by AI-Hub.

It focuses exclusively on complex clinical scenarios generated using the "Add Constraints" method from Evol-Instruct. Unlike simple fact-retrieval datasets, these examples force the model to navigate conflicting clinical limitations (e.g., comorbidities, resource constraints, legal restrictions). Each example includes a detailed Reasoning Path generated by a high-performance thinking model (qwen3-235b-a22b-thinking-2507), ensuring the model learns the logical process behind the diagnosis and treatment plan.

Note: This repository contains only the augmented data (36,343 samples) derived from the original AI-Hub source data.

Supported Tasks

• —Medical Question Answering: Answering complex clinical queries across various specialties.
• —Clinical Reasoning: Generating step-by-step logical explanations (<think> tags) before providing the final answer.
• —Domain Adaptation: Fine-tuning models for specific medical departments using the provided configs.

Dataset Structure

Data Instances

The dataset is provided in a JSONL format compatible with Chat templates. Each instance contains a conversation history where the assistant provides a hidden reasoning process within <think> tags before the final response.

{
  "messages": [
    {
      "role": "user",
      "content": "62세 여성 환자는 우측 유방암 진단 후... (중략) ... 이러한 복합적 제약 조건을 고려할 때, 림프부종 관리에 가장 적절한 설명은?"
    },
    {
      "role": "assistant",
      "content": "<think>\nHere's a thinking process that leads to the suggested answer:\n1.  **Analyze Constraints**:\n    * **Constraint 1 (Comorbidities)**: Heart failure (NYHA class III), Atrial fibrillation.\n    * **Constraint 2 (Medication)**: Rivaroxaban (anticoagulant).\n    * **Constraint 3 (History)**: Recent bacterial skin infection, Radiation-induced fibrosis.\n\n2.  **Evaluate Options based on Constraints**:\n    * *Option 1*: Incorrect. HF requires caution but is not an absolute contraindication for upper limb compression.\n    * *Option 3*: Correct. Waiting 4 weeks post-infection satisfies the infection constraint; minimizing pressure addresses the fibrosis/pain constraint.\n\n3.  **Conclusion**: Option 3 is the only viable strategy.\n</think>\n### **의료 문제 분석 및 단계별 추론 과정**\n\n---\n\n#### **1. 문제의 핵심 개념 파악**\n...\n#### **4. 최종 답에 대한 설명**\n**정답: 3**..."
    }
  ]
}

Data Fields

• —messages: A list of dictionaries representing the conversation.
• —role: The role of the speaker (user or assistant).
• —content: The text content.
• —User Content: Contains the medical question with added constraints.
• —Assistant Content: Contains the Thinking Process (<think>...</think>) followed by the Final Answer.

Dataset Statistics

This repository covers 13 Medical Domains. You can load the entire dataset or specific domains using the config_name.

Dataset Creation

1. Source Data

The original question-answer pairs were sourced from the [Specialized Medical Knowledge Data (전문 의학지식 데이터)](https://aihub.or.kr/aihubdata/data/view.do?currMenu=115&topMenu=100&&srchDataRealmCode=REALM006&aihubDataSe=data&dataSetSn=71874) on AI-Hub. This dataset was selected for its high-quality, expert-verified medical content.

2. Data Augmentation: Evol-Instruct (Add Constraints)

To bridge the gap between textbook knowledge and complex clinical reality, we applied the Evol-Instruct methodology, focusing exclusively on the "Add Constraints" technique.

Original simple questions were rewritten to include specific limitations that force the model to weigh conflicting factors:

• —Comorbidities: Adding conditions like Heart Failure, Renal Failure, or Diabetes that complicate standard treatments.
• —Medication History: Introducing drug interactions (e.g., Anticoagulants, Immunosuppressants).
• —Environmental/Resource Constraints: Scenarios with limited equipment (e.g., power outages) or lack of specialists.

3. Reasoning Path Generation

We utilized `qwen3-235b-a22b-thinking-2507` to generate high-quality reasoning paths for the augmented data. The generated reasoning follows a structured approach:

• —Constraint Satisfaction: Verifying if each option satisfies the specific constraints added during augmentation.
• —Hierarchical Elimination: Logically ruling out distractors based on priority and contraindications.

Usage

You can load the entire dataset or a specific domain using the config_name parameter.

from datasets import load_dataset

# 1. Load ALL data (Default)
dataset_all = load_dataset("SNUH-NLP/specialist-level_medical_knowledge_dataset_sft", "all")

# 2. Load Specific Domain (e.g., Surgery)
dataset_surgery = load_dataset("SNUH-NLP/specialist-level_medical_knowledge_dataset_sft", "surgery")

# 3. Load Another Domain (e.g., Medical Law)
dataset_law = load_dataset("SNUH-NLP/specialist-level_medical_knowledge_dataset_sft", "medical_law")

# Inspect an example
print(dataset_surgery['train'][0]['messages'])

Limitations & Disclaimer

• —Synthetic Nature: The examples are synthetically generated to increase complexity. While they mimic real-world scenarios, they are artificial constructs.
• —AI-Generated Reasoning: The reasoning paths are generated by an AI model. Although they are designed to be logical and accurate, they may contain hallucinations or inaccuracies.
• —Not for Clinical Use: This dataset is intended for research and training purposes only. It should not be used as a substitute for professional medical advice, diagnosis, or treatment.

Citation

If you use this dataset, please cite the following:

@misc{specialist_level_medical_knowledge_dataset_sft,
    title     = {specialist-level_medical_knowledge_dataset_sft},
    url       = {https://huggingface.co/datasets/SNUH-NLP/specialist-level_medical_knowledge_dataset_sft},
    author    = {Healthcare AI Research Institute (HARI) of Seoul National University Hospital (SNUH)},
    month     = {January},
    year      = {2026}
}