---
dataset_info:
  features:
  - name: latex
    dtype: string
  - name: image
    dtype: image
  splits:
  - name: train
    num_bytes: 28836813681
    num_examples: 3212312
  download_size: 28836813681
  dataset_size: 28836813681
configs:
- config_name: default
  data_files:
  - split: train
    path: "*.parquet"
language:
- en
tags:
- mathematics
- latex
- computer-vision
- OCR
size_categories:
- 1M<n<10M
task_categories:
- image-to-text
pretty_name: Mathematical Expressions Dataset
viewer: true
---

# Mathematical Expressions Dataset

## Dataset Description

This dataset contains images of mathematical expressions along with their corresponding LaTeX code.
**Images will automatically be displayed as thumbnails in Hugging Face's Data Studio.**

### Dataset Summary

- **Number of files**: 17 Parquet files
- **Estimated number of samples**: 3,212,312
- **Format**: Parquet optimized for Hugging Face
- **Features configured for thumbnails**: ✅
- **Columns**:
  - `latex`: LaTeX code of the mathematical expression (string)
  - `image`: Image of the mathematical expression (Image with decode=True)

### Supported Tasks

- **Image-to-Text**: Conversion of mathematical expression images to LaTeX code
- **OCR**: Optical character recognition for mathematical characters
- **Mathematical Expression Recognition**: Recognition of mathematical expressions

### Languages

The dataset contains mathematical expressions that are universal. The LaTeX code and associated metadata are primarily in English.

## Dataset Structure

### Data Fields

- `latex`: String with the LaTeX code that generates the mathematical expression.
- `image`: PIL image containing the rendered mathematical expression.
  - **Type**: `datasets.Image(decode=True)`
  - **Format**: Images are automatically decoded to PIL.Image.
  - **Thumbnails**: Automatically generated in Data Studio.

### Data Splits

| Split | Examples  |
|-------|-----------|
| train | 3,212,312 |

## Usage

```python
from datasets import load_dataset

# Load the dataset
# Make sure to replace {repo_id_placeholder} with your actual Hugging Face repository ID
# For example: "your_username/your_dataset_name"
dataset = load_dataset("ToniDO/TeXtract_padding")

# Access a sample
sample = dataset['train'][0]
image = sample['image']  # Already a PIL.Image thanks to decode=True
latex_code = sample['latex']

print(f"LaTeX: {latex_code}")
image.show()  # Display the image

# Images will also appear as thumbnails in Data Studio
```

## Visualization

Images will automatically be displayed as thumbnails in Hugging Face's Data Studio thanks to the Features configuration:

```python
from datasets import Features, Value, Image

features = Features({
    "latex": Value("string"),
    "image": Image(decode=True)  # This generates the thumbnails
})
```

## Dataset Creation

This dataset was created by converting WebDataset (.tar) files to Parquet format optimized for Hugging Face.

### Source Data

- **Original format**: WebDataset (.tar)
- **Conversion**: Using a custom Python script.
- **Optimization**: Parquet format with Snappy compression (default for `pyarrow.parquet.write_table` if not specified otherwise, actual compression depends on how your Parquet files were created).
- **Features**: Explicitly configured for automatic thumbnails.

## Technical Details

### Image Handling

- **Storage**: Images are stored as bytes within the Parquet file.
- **Decoding**: Automatic to PIL.Image when loading the dataset using `datasets.Image(decode=True)`.
- **Thumbnails**: Automatically generated by Hugging Face Data Studio.
- **Compatibility**: Works with image formats supported by PIL (Pillow) when decoded.

## Considerations for Using the Data

### Social Impact of Dataset

This dataset can be used to:
- Improve mathematical expression recognition tools.
- Develop OCR systems specialized in mathematics.
- Create accessibility tools for mathematical content.

### Licensing Information

- Apache 2.0

### Citation Information

If you use this dataset in your research, please cite it as follows:

```bibtex
@misc{ToniDO_TeXtract_parquet_2025},
  author       = {ToniDO},
  title        = {{TeXtract_parquet (Parquet Format)}},
  year         = {2025},
  publisher    = {Hugging Face},
  version      = {1.0.0},
  url          = {https://huggingface.co/datasets/ToniDO/TeXtract_padding}
}
```