---
language:
- en
pretty_name: PubMedVision-Plus
task_categories:
- image-to-text
tags:
- medical
- biomedical
- multimodal
- image-captioning
- pubmed
- huatuogpt
configs:
- config_name: core
  data_files:
  - split: train
    path: core-*.parquet
- config_name: sample
  default: true
  data_files:
  - split: train
    path: sample/sample.parquet
---

<div align="center">
<h1>PubMedVision-Plus</h1>
<p><b>A large-scale medical and biomedical image–caption dataset.</b></p>
<a href="https://huggingface.co/datasets/FreedomIntelligence/PubMedVision">🤗 PubMedVision</a>
</div>

## ⚡ Introduction

**PubMedVision-Plus** is a large-scale medical image–text dataset built from scientific literature, pairing original figures with expert-authored captions to support medical multimodal pretraining.

- **Medical knowledge at scale.** Drawing on approximately **15M PubMed papers**, the collection is refined through LLM- and VLM-based quality filtering into a **12M high-quality Core**.
- **Broad visual coverage.** From radiology and pathology to microscopy, ophthalmology, and endoscopy, the data span diverse medical specialties, with modality and anatomical-region labels retained in the browsing sample.

> **License notice:** Images and captions come from medical journal articles. Please follow the source journals’ image-use and attribution requirements. See the [PMC reuse guidelines](https://pmc.ncbi.nlm.nih.gov/about/copyright/).

## 📚 Data

Each Core record includes a unique ID, a relative image path, and its corresponding caption. The browsing sample additionally retains modality and anatomical-region labels and embeds the images.

| Data | Image–caption pairs | Construction |
| --- | ---: | --- |
| **Core** | **12,444,904** | Medical literature collection, followed by LLM/VLM quality filtering. |

> Original figures and author-written captions are preserved without generative rewriting, avoiding hallucinations introduced by synthetic captions. Filtering targets defective images and problematic image–text pairs.

## 🚀 Usage

```text
PubMedVision-Plus/
├── README.md
├── core-*.parquet       # IDs, captions, and relative image paths
├── images/images-*.tar # Original image archives
└── sample/sample.parquet # Examples with embedded images
```

The Core collection contains 12,444,904 pairs. The browsing sample currently contains 50 examples.

Load the image-embedded sample Parquet from the downloaded repository:

```python
from datasets import load_dataset

dataset = load_dataset("parquet", data_files="sample/sample.parquet", split="train")
image, caption = dataset[0]["image"], dataset[0]["caption"]
```

For the full Core release, extract all image tar shards **from the repository root** (four archives in parallel):

```bash
find images -maxdepth 1 -type f -name '*.tar' -print0 \
  | xargs -0 -P 4 -I {} tar -xf '{}' -C .
```

Extraction preserves the relative image paths in the Parquet metadata. Load `core-*.parquet` to access IDs, captions, and the corresponding image paths.

## 📄 Related work

PubMedVision-Plus is introduced in *Incentivizing Medical Vision Capabilities from Large-Scale Multimodal Pre-training* (NeurIPS 2026). If you use this dataset, please cite:

```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}
}
```