README.md

---
dataset_info:
  features:
  - name: file_name
    dtype: image
  - name: url
    dtype: string
  - name: book
    dtype: string
  - name: description
    dtype: string
  - name: context
    dtype: string
  - name: captions
    dtype: string
  splits:
  - name: full
    num_bytes: 40433558
    num_examples: 2020
  - name: train
    num_bytes: 69779282.08386138
    num_examples: 1990
  - name: test
    num_bytes: 1036459.2574257426
    num_examples: 20
  - name: validation
    num_bytes: 613312.1287128713
    num_examples: 10
  download_size: 137177467
  dataset_size: 111862611.47
configs:
- config_name: default
  data_files:
  - split: full
    path: data/full-*
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
  - split: validation
    path: data/validation-*
language:
- en
tags:
- math
- mathematics
- education
- accessibility
- STEM
pretty_name: MIDAL Dataset
size_categories:
- 1K<n<10K
license: cc-by-nc-sa-4.0
---
# Dataset Card for MIDAL Dataset

The Math Image Descriptions for Accessible Learning (MIDAL) Dataset is a curated collection of 2,020 mathematical figures paired with descriptions, captions, contextual information, and source metadata. It was created to support research in OCR-free math understanding, math image description generation, and improving accessibility in mathematics education.

> [!IMPORTANT]
> **Citation Requirement:** If you use this dataset (or any of its splits) in your research, presentations, or products, please formally cite our associated arXiv paper. See the [Citation](#citation) section below for BibTeX and APA formats.

In this repository, you will find:
* the full dataset
* the training, testing, and validation subsets used for (paper in progress)


For in-depth information regarding the dataset, please read our [paper](https://arxiv.org/abs/2608.00868).

## Dataset Details

### Dataset Description

<!-- Provide a longer summary of what this dataset is. -->
The following table depicts the split in the dataset to form the training, testing, and validation subsets used for the paper (link):

| Split       | Size | Description                     |
|-------------|------|---------------------------------|
| Full         | 2020 | The entire dataset (no splits)
| Train       | 1990 | Used for model training         |
| Validation  | 10   | Used for hyperparameter tuning  |
| Test        | 20   | Held-out evaluation set         |

#### Split Logic
The splits were created by
1. shuffle dataset with seed = 36
2. select 30 images for temporary split (seed = 36)
3. split the temporary split to produce 10 entries for validation and 20 for testing (seed = 36)
4. the remaining data forms the training subset.



### Dataset Metadata

- **Curated by:** Rebeka Popek, Vaghawan Ojha, Young Hwan You
- **Language(s) (NLP):** English
- **License:** CC BY-NC-SA 4.0
- **Paper:** [MIDAL: Math Image Descriptions for Accessible Learning](https://arxiv.org/abs/2608.00868)

## Uses
MIDAL is intended for training and evaluating models in:
* OCR-free math description generation and
* accessibility for mathematical content.


### Out-of-Scope Use
This dataset is not intended to be used commercially as that will violate the Creative Commons copyrights of many sources.


## Dataset Features

| Field       | Type     | Description                       |
|-------------|----------|-----------------------------------|
| file_name   | Image    | The image itself                  |
| description | string   | Figure description                |
| context     | string   | Additional contextual information |
| captions    | string   | Figure captions that appeared below the image  |
| book        | string   | Source book title           |
| url         | string   | Source URL         |


## Dataset Creation

### Source Data

<!-- This section describes the source data (e.g. news text and headlines, social media posts, translated sentences, ...). -->
The data was collected only from open educational resources (OER). If using this dataset, please respect their intellectual property and 
copyrights.

### Data Collection and Processing

<!-- This section describes the data collection and processing process such as data selection criteria, filtering and normalization methods, tools and libraries used, etc. -->

The dataset was produced using 
* Python's Playwright library to collect the metadata
* Label Studio for annotation and
* manual curation to ensure quality and accuracy.

Quality checks included:
* rewriting image descriptions to fit NWEA's image description guidelines
* converting all math content into LaTeX (in amsmath formatting) and
* adding white backgrounds to images with transparent backgrounds.

### Maintainer
Huggingface: @rpopek

email contact: rebekapopek@gmail.com


## Citation

<!-- If there is a paper or blog post introducing the dataset, the APA and Bibtex information for that should go in this section. -->

**BibTeX:**

```bibtex
@misc{popek2026midalmathimagedescriptions,
      title={MIDAL: A Dataset of Math Image Descriptions for Accessible Learning}, 
      author={Rebeka Popek and Vaghawan Ojha and Young Hwan You},
      year={2026},
      eprint={2608.00868},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2608.00868}, 
}
```