---
configs:
- config_name: default
  data_files:
  - split: train
    path: "images.zip"
task_categories:
- image-captioning
- image-text-retrieval
language:
- en
tags:
- mathematics
- multimodal
- benchmark
- wikipedia
size_categories:
- 1K<n<10K
---

# MAT-VB: MAthematical Text-Vision Benchmark

🏆 **Best Resource Paper Award** — *2025 ACM/IEEE Joint Conference on Digital Libraries (JCDL)*

Official Hugging Face dataset repository for **MAT-VB**, a benchmark built on Wikipedia math images designed to advance research in multi-modal mathematical tasks, including visual-text retrieval and math image captioning.

---

## 🔗 Links

* **GitHub Repository:** [AIIRLab/MATVB](https://github.com/AIIRLab/MATVB)
* **Paper (IEEE Xplore):** [MAT-VB: MAthematical Text-Vision Benchmark](https://ieeexplore.ieee.org/abstract/document/11363711)

---

## 📁 Dataset Contents

MAT-VB provides a curated collection of mathematical visual data from Wikipedia for evaluating multimodal AI models:

* **`images.zip`**: Contains all ~4,000 Wikipedia math image files (216 MB uncompressed archive).
* **Metadata & Annotations**: Includes descriptive captions, query-image mappings, and split indexes for training and evaluation pipelines.

---

## 🚀 Usage with Python `datasets`

Hugging Face automatically extracts and parses `images.zip` when loading through the `datasets` library:

```python
from datasets import load_dataset

# Load the dataset directly
dataset = load_dataset("AIIRLab/MATVB")

# Access a sample
example = dataset["train"][0]
image = example["image"]
caption = example.get("caption", "")