𝑩𝑰𝑮 𝑵𝑬𝑾𝑺‼️

📮 [2025-08] We have updated to a larger dataset, which contains nearly 80 million samples, with significant improvements in both data quality and diversity. [link]

📮 [2024-02] We trained a formula recognition model, 𝐓𝐞𝐱𝐓𝐞𝐥𝐥𝐞𝐫, using the latex-formulas dataset. It can convert LaTeX formulas into images and boasts high accuracy and strong generalization capabilities, covering most formula recognition scenarios.

For more details, please refer to the 𝐓𝐞𝐱𝐓𝐞𝐥𝐥𝐞𝐫 GitHub repository.

Dataset Description

中文版本

There are two datasets: raw_formulas and cleaned_formulas(This dataset has 550K formula-image pairs).

We scraped approximately 1 million LaTeX formula image-text pairs from arxiv that were uncleaned and without text segmentation to create the raw_formulas dataset. After cleaning the raw_formulas dataset and integrating it with the im2latex-100K dataset, we obtained the cleaned_formulas dataset, which has 550K formula-image pairs.

To render the images corresponding to the formulas, the following external packages are needed:

• amsmath
• amsfonts
• amssymb
• mathtools

Usage

for raw_formulas dataset:

from datasets import load_dataset
data = load_dataset("OleehyO/latex-formulas", "raw_formulas") 

for cleaned_formulas dataset:

from datasets import load_dataset
data = load_dataset("OleehyO/latex-formulas", "cleaned_formulas") 

Details About the raw_formulas Dataset

We scraped LaTeX formulas containing the following environments:

• equation
• align
• align*
• gather
• gather*

The formulas do not include the following content:

• \label
• %
• \quad
• \qquad
• \vspace
• \hspace
• \resizebox
• \scalebox
• \rotatebox
• \parbox
• \fbox
• \makebox
• \raisebox
• \addvspace
• \hfill
• \vfill
• \textwidth
• \textheight
• \rule

Preprocessing Details of the cleaned_formulas Dataset

Cleaning

• We removed some useless junk data from both raw_formulas and im2latex-100K.
• We deleted overly complex formulas from both raw_formulas and im2latex-100K:
  • Formulas were deleted if the aspect ratio of the corresponding rendered image was greater than 0.8.
  • Formulas with a character length greater than 200 were deleted.
• In the formulas from both raw_formulas and im2latex-100K, the following content was removed:
  • \tag
  • \text
  • \begin{split}
  • \end{split}
  • \nonumber
  • \notag
• The equation, equation*, align, \[...\] environments in raw_formulas were all replaced with the align* environment.
• We deleted formulas from raw_formulas that contained custom macros.