dataset_info:
features:
- name: image
dtype: image
- name: extracted_text
dtype: string
splits:
- name: train
num_bytes: 1319873205
num_examples: 5349
download_size: 1295312059
dataset_size: 1319873205
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
license: cc-by-nc-sa-4.0
language:
- or
size_categories:
- 1K<n<10K
Odia OCR Synthetic Dataset
Dataset Summary
The Odia OCR Synthetic Dataset is a curated collection of synthetically generated Odia text images, created specifically for OCR training, and paired with accurate Odia Unicode transcriptions.
This dataset is designed to support:
- Training and fine-tuning Odia OCR models
- Vision-language model (VLM)–based OCR training
- Data augmentation for low-resource Odia script recognition
- Improving robustness across fonts, layouts, and visual variations
The synthetic nature of the dataset allows controlled variation in font styles, spacing, backgrounds, and noise patterns, helping OCR models generalize better to real-world scanned and photographed documents.
Dataset Structure
Features
Each sample in the dataset contains the following fields:
image(image)
A synthetically generated image containing Odia text, rendered with diverse fonts, sizes, spacing, and background variations.extracted_text(string)
The ground-truth Odia Unicode transcription corresponding to the text rendered in the image.
Citation
If you use the Odia OCR Synthetic Dataset in your research, experiments, or applications, please cite it as follows:
BibTeX
@dataset{sahu_odia_ocr_synthetic_2026,
title = {Odia OCR Synthetic Dataset},
author = {Sahu, Siba Prasad and OdiaGenAI OCR Team},
year = {2026},
publisher = {Hugging Face},
url = {https://huggingface.co/datasets/OdiaGenAIOCR/odia-ocr-merged}
}