---
language:
- en
task_categories:
- visual-question-answering
size_categories:
- n<1K
tags:
- medical
- multimodal
- evaluation
- rubric
- huatuogpt-vision-2
configs:
- config_name: default
  data_files:
  - split: test
    path: medmm_expert_rubric.json
---

<div align="center">
<h1>MedMM-ExpertRubric</h1>
<a href="https://openreview.net/pdf?id=0BXKsjKlFJ">📄 Paper</a> |
<a href="https://github.com/FreedomIntelligence/HuatuoGPT-Vision-2">💻 GitHub</a>
</div>

## ⚡ Introduction

**MedMM-ExpertRubric** is an open-ended medical multimodal benchmark with **physician-verified rubrics**, released with **HuatuoGPT-Vision-2**. It contains **500 questions, 538 images, and 11,088 rubric items**.

- 👨‍⚕️ **Physician-verified.** Licensed physicians filter every rubric item, keeping only core image-understanding points.
- 🔬 **Fine-grained scoring.** About 22 weighted rubric items per question. Positive items reward key findings, and negative items penalize medical errors.
- 🩻 **Diverse and held-out.** Covers radiology, pathology, ultrasound, endoscopy, and more, with 32 multi-image questions. None of the cases is used in HuatuoGPT-Vision-2 training.
- 🤝 **Aligned with physicians.** Rankings agree with physician pairwise preference, making it a low-cost proxy for expert evaluation.

## 📚 What's inside

| File | Content |
| --- | --- |
| `inference_inputs.json` | Question and images only. Use this as model input. |
| `medmm_expert_rubric.json` | Full records with rubrics, caption, modality, and anatomy. For scoring only. |

## 🚀 Usage

Load the downloaded data:

```python
import json
from pathlib import Path
from PIL import Image

root = Path("/path/to/MedMM-ExpertRubric")
dataset = json.loads((root / "inference_inputs.json").read_text())
case = dataset[0]
images = [Image.open(root / path) for path in case["images"]]
print(case["question"])
```

Provide the question and images in order to the model. Use the rubrics for evaluation only. Do not include the caption in model input.

An LLM grader marks each rubric item as met or not met. The score is the sum of points for met items divided by the total positive points. We use GPT-4.1 in the paper, and other strong LLMs can also serve as the grader.

## 📖 Citation

```bibtex
@inproceedings{chen2026incentivizing,
  title={Incentivizing Medical Vision Capabilities from Large-Scale Multimodal Pre-training},
  author={Chen, Junying and Cai, Zhenyang and Yang, Yunjin and Cai, Kunyuan and Wang, Rongsheng and Wang, Xidong and Yu, Geyang and Feng, Xiangyi and Wang, Benyou},
  booktitle={The Fortieth Annual Conference on Neural Information Processing Systems},
  year={2026}
}
```