---
license: openrail
task_categories:
- image-to-text
tags:
- ocr
- latex-ocr
- image2latex
dataset_info:
- config_name: cleaned_formulas
  features:
  - name: image
    dtype: image
  - name: latex_formula
    dtype: string
  splits:
  - name: train
    num_bytes: 2918992848.46
    num_examples: 552340
  download_size: 2778067493
  dataset_size: 2918992848.46
- config_name: raw_formulas
  features:
  - name: latex_formula
    dtype: string
  splits:
  - name: train
    num_bytes: 240965616
    num_examples: 1006245
  download_size: 89507618
  dataset_size: 240965616
configs:
- config_name: cleaned_formulas
  data_files:
  - split: train
    path: cleaned_formulas/train-*
- config_name: raw_formulas
  data_files:
  - split: train
    path: raw_formulas/train-*
---

# 𝑩𝑰𝑮 𝑵𝑬𝑾𝑺‼️

📮 [2𝟎2𝟒-𝟎2] We trained a formula recognition model, [𝐓𝐞𝐱𝐓𝐞𝐥𝐥𝐞𝐫](https://github.com/OleehyO/TexTeller?tab=readme-ov-file), 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](https://github.com/OleehyO/TexTeller?tab=readme-ov-file).

# Dataset Description
> [中文版本](./README_zh.md)

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](https://zenodo.org/records/56198#.V2px0jXT6eA) 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:

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

for **cleaned_formulas** dataset:
```python
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](https://zenodo.org/records/56198#.V2px0jXT6eA).
* We deleted overly complex formulas from both *raw_formulas* and [im2latex-100K](https://zenodo.org/records/56198#.V2px0jXT6eA):
  * 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](https://zenodo.org/records/56198#.V2px0jXT6eA), 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.