---
dataset_info:
  features:
  - name: image
    dtype: image
  - name: latex
    dtype: string
  - name: latex_norm
    dtype: string
  - name: page_id
    dtype: string
  - name: expr_type
    dtype: string
  splits:
  - name: train
    num_bytes: 222324499.18
    num_examples: 103940
  - name: val
    num_bytes: 41861369.289
    num_examples: 18371
  - name: test
    num_bytes: 94309716.225
    num_examples: 44383
  download_size: 340062142
  dataset_size: 358495584.694
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: val
    path: data/val-*
  - split: test
    path: data/test-*
task_categories:
- image-to-text
tags:
- latex
- math
- ocr
size_categories:
- 100K<n<1M
---
# IBEM Cropped: Bounding Box Mathematical Expressions from Scientific Pages

A processed version of the **IBEM Dataset** (Anitei et al., 2023), designed for the training and evaluation of image-to-markup models, expression recognition, and related document understanding tasks. Each sample in this dataset consists of a cropped image region containing a **single mathematical expression** (either displayed or embedded), along with its corresponding LaTeX annotations.

## Dataset Summary

This dataset is derived from the original **IBEM dataset** by extracting individual mathematical expressions using the provided bounding boxes, and cropping them from their original document pages. All page-level coordinates were specified as percentages and converted to absolute pixel coordinates during cropping.

Each record includes:
- A cropped image of the mathematical expression.
- The original LaTeX string (`latex`).
- A normalized LaTeX string (`latex_norm`) with visual formatting removed.
- The ID of the original page.
- The expression type (`isolated` or `embedded`).

## Dataset Structure

**Features:**

```python
features = Features({
    "image": Image(),
    "latex": Value("string"),
    "latex_norm": Value("string"),
    "page_id": Value("string"),
    "expr_type": Value("string"),
})
````

**Splits:**

The dataset is split into standard `train`, `val`, and `test` subsets based on the official partition lists provided by the authors (`Tr*.lst`, `Va*.lst`, `Ts*.lst`).

## Source

**Citation:**

> Anitei, D., Sánchez, J. A., & Benedí, J. M. (2023). The IBEM Dataset: a large printed scientific image dataset for indexing and searching mathematical expressions (1.0) \[Data set]. Zenodo. [https://doi.org/10.5281/zenodo.7963703](https://doi.org/10.5281/zenodo.7963703)

## Original Dataset Description

The IBEM dataset consists of 600 LaTeX documents with a total of 8,272 pages, containing:

* **29,603 displayed mathematical expressions**
* **137,089 embedded mathematical expressions**

These were extracted from the LaTeX source files of documents in the **KDD Cup Collection**. The dataset supports a variety of tasks, including:

* Mathematical expression detection and extraction
* LaTeX recognition from printed page images
* Indexing and search of STEM content in large-scale document archives

## Preprocessing Notes

* **Preamble fields** containing LaTeX macros were excluded.
* All expressions were **cropped** to their respective bounding boxes and stored as image snippets.
* Only the `latex` and `latex_norm` fields were retained for expression text.
* Expressions split across multiple lines were treated as-is, using the duplicated complete LaTeX string provided by the original dataset.

## Licensing

Distributed under the same license as the original IBEM dataset.
Creative Commons Attribution 4.0 International