---
license: mit
dataset_info:
  features:
  - name: image
    dtype: image
  - name: text
    dtype: string
  splits:
  - name: train_shard_034
    num_bytes: 3070462057.0
    num_examples: 5000
  download_size: 3070433781
  dataset_size: 3070462057.0
pretty_name: tamily 1
language:
- ta
source_datasets:
- sasicodes/solvari-1
task_categories:
- image-to-text
tags:
- Vaṭṭeḻuttu
configs:
- config_name: default
  data_files:
  - split: train_shard_034
    path: data/train_shard_034-*
---


# Tamily-1: Ancient Tamil OCR Synthetic Dataset

## Description

- **Repository:** [sasicodes/tamily-1](https://huggingface.co/datasets/sasicodes/tamily-1)
- **Point of Contact:** [@sasicodes](https://huggingface.co/sasicodes)

### Summary

Tamily-1 is an ancient Tamil OCR synthetic dataset generated from the first 200,000 rows of [Solvari-1](https://huggingface.co/datasets/sasicodes/solvari-1), a large Tamil text corpus. The dataset contains rendered images of Tamil text with various augmentations and styles, making it suitable for training OCR models.

### Fields

- `image`: PNG image of rendered Tamil text
- `text`: Original Tamil text

### Data Splits

The dataset is split into shards of 5,000 samples each, named as `train_shard_XXX`.

#### Annotation process

Each text is rendered with:
- Random paper style (Palm Leaf, Pale Palm Leaf, Red Stone, White Stone, Paper)
- Random background style (No Lines, With Lines, Blurred, With Lines and Noise)
- Random augmentation (Rotation, Perspective, Stain, Ink Bleed)

### License

MIT License


```bibtex
@misc{tamily-1,
  author = {sasicodes},
  title = {Tamily-1: Ancient Tamil OCR Synthetic Dataset},
  year = {2025},
  publisher = {Hugging Face},
  journal = {Hugging Face Hub},
  howpublished = {\url{https://huggingface.co/datasets/sasicodes/tamily-1}}
}
```