README.md

metadata

license: mit
language:
  - am
size_categories:
  - 100K<n<1M

Fidel: A Large-Scale Sentence Level Amharic OCR Dataset

Overview

Fidel is a comprehensive dataset for Amharic Optical Character Recognition (OCR) at the sentence level. It contains a diverse collection of Amharic text images spanning handwritten, typed, and synthetic sources. This dataset aims to advance language technology for Amharic, serving critical applications such as digital ID initiatives, document digitization, and automated form processing in Ethiopia.

Citation

If you use the Fidel: A Large-Scale Sentence Level Amharic OCR Dataset in your work, please cite the associated pre-print using the following BibTeX entry:

@article{Fidel2025,
  author = {Chamisso, Tunga Tessema and Guda, Blessed and Adego, Bereket Retta and Sagbo, Carmel Prosper and Ashungafac, Gabrial Zencha and Gueye, Assane},
  title = {{Fidel: A Large-Scale Sentence Level Amharic OCR Dataset}},
  journal = {Research Square},
  year = {2025},
  doi = {10.21203/rs.3.rs-8118465/v1},
  url = {[https://doi.org/10.21203/rs.3.rs-8118465/v1](https://doi.org/10.21203/rs.3.rs-8118465/v1)}
}

Dataset Structure

The dataset is organized into train and test splits:

fidel-dataset/
├── train/   # training images (handwritten, typed, and synthetic)
├── test/    # test images (handwritten, typed, and synthetic)
└── metadata.json  # Croissant metadata file
├── train_labels.json # Contains filenames and corresponding text labels
└── test_labels.json  # Contains filenames and corresponding text labels

Labels Format

Each CSV file contains the following columns:

• image_filename: Name of the image file
• line_text: The Amharic text content in the image
• type: The source type (handwritten, typed, or synthetic)
• writer: The writer number (for handwritten types only)

Example Labels

image_filename	text
25_line_4.png	ዲግሪዎች የህዝብ አስተዳደር ትምህርት ተምረው አንዳገኟቸው ሲገልጹ የቆዩ ሲሆን ይህንንም በፓርላማ ድረገጽ ፣ በፌስቡክ ፣ ዊኪፔድያ ፣
3_line_2.png	ዮርክ ኬኔዲ አየር ጣቢያ ተነስቶ ሎንዶን ሂዝሮው አየር ጣቢያ አረፈ። ዝምባብዌም በመንግስት ለታገዘ ዝርፊያ እንዲሁም ለድህነትና በሽታ እጇን

Usage

# Install git-lfs if not already installed
git lfs install

# Clone with LFS support for large files
git clone https://huggingface.co/datasets/upanzi/fidel-dataset
cd fidel-dataset

# Pull LFS files (zip archives)
git lfs pull

# Extract the archives
unzip train.zip
unzip test.zip

Dataset Statistics

Overall Statistics

• Total samples: 193,185
• Training samples: 169,499
• Test samples: 23,686

By Source Type

• Handwritten: 40,265 samples
• Typed: 28,342 samples
• Synthetic: 61,530 samples
• HDD: 63,048 samples

Image Characteristics

• Average image width: varies by type (handwritten: 2,480px, typed: 2,482px, synthetic: 2,956px)
• Average image height: varies by type (handwritten: 199px, typed: 71px, synthetic: 244px)
• Average aspect ratio: varies by type (handwritten: 14.0, typed: 19.5, synthetic: 11.6)

Text Characteristics

• Average text length: varies by type (handwritten: 62.0 characters, typed: 95.2 characters, synthetic: 74.7 characters)
• Average word count: varies by type (handwritten: 11.3 words, typed: 16.9 words, synthetic: 14.7 words)
• Unique characters: 249 in handwritten, 200 in typed, 190 in synthetic

License

This dataset is released under: MIT License

Acknowledgments

We thank all contributors who provided handwritten samples and the organizations that supported this data collection effort.