---
language:
- en
task_categories:
- visual-question-answering
size_categories:
- 10K<n<100K
tags:
- medical
- multimodal
- reinforcement-learning
- huatuogpt-vision-2
configs:
- config_name: open_ended
  data_files:
  - split: train
    path: data/open_ended/*.jsonl
- config_name: multiple_choice
  data_files:
  - split: train
    path: data/multiple_choice/*.jsonl
---

<div align="center">
<h1>PubMedVision-RL</h1>
<a href="https://openreview.net/pdf?id=0BXKsjKlFJ">📄 Paper</a> |
<a href="https://github.com/FreedomIntelligence/HuatuoGPT-Vision-2">💻 GitHub</a> |
<a href="https://huggingface.co/datasets/FreedomIntelligence/PubMedVision-Plus">🤗 PubMedVision-Plus</a> |
<a href="https://huggingface.co/datasets/FreedomIntelligence/PubMedVision-RL">🤗 PubMedVision-RL</a>
</div>

## ⚡ Introduction

**PubMedVision-RL** is a 70K medical multimodal RL dataset for **eliciting medical vision capabilities** in VLMs. It is the RL data behind **HuatuoGPT-Vision-2**.

- 🩺 **Built from real medical data.** Every question comes from a real PubMed image-caption pair in [PubMedVision-Plus](https://huggingface.co/datasets/FreedomIntelligence/PubMedVision-Plus), grounded in expert-written captions.
- 🎯 **Two complementary types.** Multiple-choice questions with verifiable labels, and open-ended questions with weighted rubrics.
- ✅ **Quality filtered.** A second GPT-5.4 pass validates every question and rubric. Ambiguous items and those going beyond the caption are removed.
- 🔥 **Difficulty filtered.** Questions that Qwen3-VL-8B-Instruct already solves are dropped, so each sample gives a useful RL signal.

## 📚 What's inside

| Configuration | Samples | Reward |
| --- | ---: | --- |
| `multiple_choice` | 35,331 | Verifiable (option label) |
| `open_ended` | 35,154 | Weighted rubrics |

Both configurations use the `train` split and include the question, image paths, and source caption.

## 🚀 Usage

After downloading the repository:

```python
from datasets import load_dataset

root = "/path/to/PubMedVision-RL"
open_ended = load_dataset(root, "open_ended", split="train")
multiple_choice = load_dataset(root, "multiple_choice", split="train")
print(open_ended[0])
```

Image paths are relative to the repository root. Provide the question and images to the model. Use rubrics or answer labels to compute rewards. Captions are retained for reference.

## 📖 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}
}
```