README.md

---
license: mit
tags:
  - img2latex
  - latex-ocr
  - handwritten mathematical expressions
  - printed mathematical expressions
size_categories:
  - 1M<n<10M
dataset_info:
  config_name: default
  features:
    - name: "__key__"
      dtype: string
    - name: ".png"  # Or other image extensions like .jpg, depending on your dataset
      dtype: image   # Or binary if not auto-decoded by the `datasets` library
    - name: ".tex"
      dtype: string  # Will be bytes initially, requires decoding
  splits:
    - name: train
      num_bytes: <INSERT_TOTAL_DATASET_SIZE_IN_BYTES>
      num_examples: 3200000
---

# TeXtract_dataset (WebDataset Format)

This repository contains approximately **3.2 million** pairs of mathematical expression images and their corresponding LaTeX source code, packaged in **WebDataset** format for large-scale training.

The dataset is based on and derived from the original [hoang-quoc-trung/fusion-image-to-latex-datasets](https://huggingface.co/datasets/hoang-quoc-trung/fusion-image-to-latex-datasets), transformed for more efficient access.

---

## 📂 Dataset Structure

Each WebDataset shard (`.tar`) contains multiple samples. Each sample groups files sharing a common identifier (`__key__`):

* `__key__` (string): Unique sample ID (e.g., `sample_000000123`).
* Image file (`.png`, `.jpg`, etc.): Binary data of the mathematical expression.
* `.tex`: UTF-8 text file with the corresponding LaTeX code.
* `__url__` (string): URL or path to the source shard (automatically added).

```
shard-000000.tar
├── sample_000000000.png
├── sample_000000000.tex
├── sample_000000001.png
├── sample_000000001.tex
└── ...
```

> **Note:** When browsing in Hugging Face Data Studio:
>
> * Image metadata (dimensions) may be shown instead of the actual content.
> * `.tex` files may appear Base64-encoded. This is only a preview; the underlying data is UTF-8.

---

## 🚀 How to Use

### 1. Using the `datasets` library (recommended)

```python
from datasets import load_dataset
from PIL import Image
import io

DATASET_ID = "ToniDO/TeXtract_dataset"

try:
    ds = load_dataset(DATASET_ID, split="train", trust_remote_code=True)
except ValueError:
    ds = load_dataset(DATASET_ID, trust_remote_code=True)

for i, sample in enumerate(ds):
    print(f"Sample {i}: {sample['__key__']}")

    # Load image
    for ext in ['.png', '.jpg', '.jpeg']:
        if ext in sample:
            img_data = sample[ext]
            img = (
                img_data
                if isinstance(img_data, Image.Image)
                else Image.open(io.BytesIO(img_data if isinstance(img_data, bytes) else img_data['bytes']))
            )
            print(f"Image ({ext}), size: {img.size}")
            break

    # Decode LaTeX
    tex_bytes = sample.get('.tex')
    if isinstance(tex_bytes, (bytes, bytearray)):
        latex = tex_bytes.decode('utf-8')
        print(latex[:100])

    if i >= 2:
        break
```

### 2. Using the `webdataset` library

```python
import webdataset as wds
from PIL import Image
import io

urls = "path/to/shards/math_dataset-{000000..000349}.tar"

dataset = (
    wds.WebDataset(urls)
       .decode(
           wds.handle_extension("pil", "png"),
           wds.handle_extension("pil", "jpg"),
           handler=wds.ignore_and_continue
       )
)

for i, sample in enumerate(dataset):
    print(f"Sample {i}: {sample['__key__']}")

    # Image
    img = None
    for ext in ["png", "jpg", "jpeg"]:
        if ext in sample and isinstance(sample[ext], Image.Image):
            img = sample[ext]
            break
    if img:
        print(f"Size: {img.size}")

    # LaTeX
    tex = sample.get('.tex')
    if isinstance(tex, (bytes, bytearray)):
        print(tex.decode('utf-8')[:100])

    if i >= 2:
        break
```

> **Training tips:**
>
> * Decode LaTeX from UTF-8.
> * Preprocess images (resize, normalize, augment).
> * Tokenize LaTeX code according to your vocabulary.
> * Shuffle shards and samples for effective training.

---

## File Types
```console
  .bmp
  .dvi
  .jpg
  .png

```

## 📖 Citation

If you use this dataset, please cite the original work:

```bibtex
@misc{hoang2024fusion,
  author       = {Hoang, Quoc Trung},
  title        = {Fusion Image-to-LaTeX Datasets},
  year         = {2024},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/hoang-quoc-trung/fusion-image-to-latex-datasets}
}
```

And to reference this WebDataset version:

```bibtex
@misc{ToniDO_TeXtract_webdataset_2025,
  author       = {ToniDO},
  title        = {{TeXtract_dataset (WebDataset Format)}},
  year         = {2025},
  publisher    = {Hugging Face},
  version      = {1.0.0},
  url          = {https://huggingface.co/datasets/ToniDO/TeXtract_dataset}
}
```

---

## 📝 Authors

* ToniDO

---

## 📜 License

This project is licensed under the **MIT License**. See the [LICENSE](LICENSE) file for details.