Medical-O1-Reasoning

Description

Medical-O1-Reasoning is an environment for evaluating medical reasoning capabilities. It contains 90,120 medical questions from the HuatuoGPT-o1 dataset covering clinical diagnosis, treatment, pathophysiology, procedures, pharmacology, and genetics in both English and Chinese.

Capabilities

• Complex medical reasoning
• Clinical diagnosis and treatment planning
• Medical knowledge across multiple specialties
• Bilingual medical question answering (English and Chinese)

Compute Requirements

Agents are given a standard environment with no sandbox or file system access.

License

Apache 2.0.

Tasks

There are four splits in this environment:

• en: 19,704 tasks (English medical questions)
• en_mix: 24,887 tasks (English medical + general instruction)
• zh: 20,171 tasks (Chinese medical questions)
• zh_mix: 25,358 tasks (Chinese medical + general instruction)

Reward Structure

This is a single-turn environment. The agent replies with its medical answer as an ordinary message (no tool call), which ends the rollout. An LLM grader (gpt-5-mini) evaluates the response against expert reference answers on three criteria: (1) whether the answer captures the key medical concepts from the reference, (2) whether the reasoning is clinically sound, and (3) whether the answer is free of factual errors or critical omissions. Reward is binary: 1.0 if correct, 0.0 if incorrect.

Data

Data consists of four Parquet files sourced from HuggingFace FreedomIntelligence/medical-o1-reasoning-SFT. Each row contains a medical question and expert reference answer. Data is stored on the OpenReward platform.

Tools

None. The model is given no tools: it answers the medical question as an
ordinary message, and that message ends the rollout.

Grading runs through a hidden @terminal tool, which sends the message text
to the LLM grader (gpt-5-mini) alongside the expert reference answer.

Time Horizon

Single-turn. The agent reads the medical question and replies once with a plain message.

Environment Difficulty

[Put environment difficulty statistics here]

Other Environment Requirements

OpenAI API key required for LLM-based grading. Pass via secrets={"openai_api_key": "..."}.

Safety

Agents in Medical-O1-Reasoning answer medical questions in a standard environment. Models trained on this data should not be relied upon for medical advice.

Citation

@article{chen2024huatuogpto1,
  title={HuatuoGPT-o1, Towards Medical Complex Reasoning with LLMs},
  author={Chen, Junying and Cai, Zhenyang and Ji, Ke and Wang, Xidong and Liu, Wanlong and Wang, Rongsheng and Hou, Jianye and Wang, Benyou},
  journal={arXiv preprint arXiv:2412.18925},
  year={2024}
}