---
license: apache-2.0
---

# MedReason: Eliciting Factual Medical Reasoning Steps in LLMs via Knowledge Graphs

<p align="center">
📃 <a href="https://arxiv.org/abs/2504.00993" target="_blank">Paper</a> ｜🤗 <a href="https://huggingface.co/UCSC-VLAA/MedReason-8B" target="_blank">MedReason-8B</a> | 📚 <a href="https://huggingface.co/datasets/UCSC-VLAA/MedReason" target="_blank">MedReason Data</a>
</p>


## ⚡Introduction

**MedReason** is a large-scale high-quality medical reasoning dataset designed to enable faithful and explainable medical problem-solving in large language models (LLMs).

- We utilize a structured medical knowledge graph (KG) to convert clinical QA pairs into logical chains of reasoning, or “thinking paths”.
- Our pipeline generates detailed reasoning for various medical questions from 7 medical datasets, resulting in a dataset of **32,682** question-answer pairs, each with detailed, step-by-step explanations. 
- By finetuning with proposed [MedReason dataset](https://huggingface.co/datasets/UCSC-VLAA/MedReason), our best model [MedReason-8B](https://huggingface.co/UCSC-VLAA/MedReason-8B), achieves *state-of-the-art* performance.

We open-sourced our model here.

## 👨‍⚕️ Model

- **Model Access**

| Model             | Base Model                                                   | Link                                                       |
| ----------------- | ------------------------------------------------------------ | ---------------------------------------------------------- |
| MedReason-8B      | [HuatuoGPT-o1-8B](https://huggingface.co/FreedomIntelligence/HuatuoGPT-o1-8B) | [Link](https://huggingface.co/UCSC-VLAA/MedReason-8B)      |
| MedReason-Llama   | [Llama-3.1-8B-Instruct](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct) | [Link](https://huggingface.co/UCSC-VLAA/MedReason-Llama)   |
| MedReason-Mistral | [Mistral-7B-Instruct-v0.2](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2) | [Link](https://huggingface.co/UCSC-VLAA/MedReason-Mistral) |

- **Deploy**: we provide a example code for direct inference with MedReason-8B. 

  Also, MedReason-8B can be deployed with tools like [vllm](https://github.com/vllm-project/vllm) or [Sglang](https://github.com/sgl-project/sglang), we provide code for model deployment using Sglang in `./src/evaluation/eval.py`

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained('UCSC-VLAA/MedReason-8B',torch_dtype="auto",device_map="auto", use_safetensors= True)
model.eval()

tokenizer = AutoTokenizer.from_pretrained('UCSC-VLAA/MedReason-8B', trust_remote_code=True, padding_side='left')

input_text = "How to stop a cough?"
messages = [{"role": "user", "content": input_text}]

inputs = tokenizer(tokenizer.apply_chat_template(messages, tokenize=False,add_generation_prompt=True), return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=2048)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

## 🙏🏼 Acknowledgement

We gratefully acknowledge the inspiring work of [HuatuoGPT-o1](https://github.com/FreedomIntelligence/HuatuoGPT-o1), which laid important groundwork for this research. We also thank the developers of the excellent tools [curator](https://github.com/bespokelabsai/curator/), [trl](https://github.com/huggingface/trl), and [sglang](https://github.com/sgl-project/sglang) for making this work possible.

## 📖 Citation

```
@misc{wu2025medreasonelicitingfactualmedical,
      title={MedReason: Eliciting Factual Medical Reasoning Steps in LLMs via Knowledge Graphs}, 
      author={Juncheng Wu and Wenlong Deng and Xingxuan Li and Sheng Liu and Taomian Mi and Yifan Peng and Ziyang Xu and Yi Liu and Hyunjin Cho and Chang-In Choi and Yihan Cao and Hui Ren and Xiang Li and Xiaoxiao Li and Yuyin Zhou},
      year={2025},
      eprint={2504.00993},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2504.00993}, 
}
```