metadata

dataset_info:
  features:
    - name: image
      dtype: image
    - name: page_id
      dtype: string
    - name: expressions
      sequence:
        - name: latex
          dtype: string
        - name: latex_norm
          dtype: string
        - name: latex_expand
          dtype: string
        - name: x_min
          dtype: float32
        - name: y_min
          dtype: float32
        - name: x_max
          dtype: float32
        - name: y_max
          dtype: float32
        - name: expr_type
          dtype: string
  splits:
    - name: train
      num_bytes: 1635422184.212
      num_examples: 4649
    - name: val
      num_bytes: 308708516
      num_examples: 856
    - name: test
      num_bytes: 709594994.982
      num_examples: 1942
  download_size: 2615759043
  dataset_size: 2653725695.194
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: val
        path: data/val-*
      - split: test
        path: data/test-*
task_categories:
  - object-detection
tags:
  - math
  - equation
  - latex
size_categories:
  - 1K<n<10K

IBEM Full: Mathematical Expression Annotations on Scientific Pages

A structured version of the IBEM Dataset (Anitei et al., 2023), designed for training and evaluation of mathematical expression detection models. This variant retains full page images and provides bounding box annotations for all mathematical expressions, enabling end-to-end layout analysis and object detection tasks.

Dataset Summary

This dataset is derived from the original IBEM corpus by preserving full-page scientific document images and annotating all mathematical expressions using their corresponding bounding boxes and LaTeX transcriptions. It is intended for:

Mathematical expression detection (localization and classification)
Document understanding and layout parsing
End-to-end systems for recognizing math expressions in context

Each record in the dataset includes:

A full-page image from a LaTeX source document
A unique page identifier
A list of expressions (expressions), where each item contains:
- The raw LaTeX string (latex)
- A normalized version of the LaTeX string (latex_norm)
- An expanded version with macros resolved (latex_expand)
- Normalized bounding box coordinates (x_min, y_min, x_max, y_max) relative to image dimensions
- The expression type (expr_type: either "isolated" or "embedded")

Dataset Structure

Features

features = Features({
    "image": Image(),
    "page_id": Value("string"),
    "expressions": Sequence({
        "latex": Value("string"),
        "latex_norm": Value("string"),
        "latex_expand": Value("string"),
        "x_min": Value("float32"),
        "y_min": Value("float32"),
        "x_max": Value("float32"),
        "y_max": Value("float32"),
        "expr_type": Value("string"),
    }),
})

Splits

The dataset is split into train, val, and test subsets using the official partition files (Tr*.lst, Va*.lst, Ts*.lst) provided by the original authors.

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

Original Dataset Description

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

29,603 isolated (displayed) mathematical expressions
137,089 embedded (in-line) mathematical expressions

It was created by parsing LaTeX source files from the KDD Cup Collection and supports a wide range of tasks such as:

Mathematical expression detection and extraction
LaTeX OCR / recognition
Search and indexing in scientific literature

Preprocessing Notes

Bounding boxes are provided in relative coordinates (percentage of image width/height).
No cropping was applied — images retain their full-page content.
Multiple expressions per page are stored in a single list under "expressions".
Expression-level latex, latex_expand, and latex_norm were directly taken from the original ground truth.
Expressions split across lines are preserved with their full content.

Licensing

This dataset is distributed under the same license as the original IBEM dataset: Creative Commons Attribution 4.0 International (CC BY 4.0)