Datasets:

Tamazight
/

Tifinagh-OCR-39K

language:
  - zgh
  - tzm
  - shi
  - gha
  - shy
  - mzb
task_categories:
  - image-to-text
tags:
  - ocr
  - tifinagh
  - tamazight
  - document-understanding
  - vision-language
pretty_name: Tifinagh OCR 39k
size_categories:
  - 10K<n<100K
license: apache-2.0

Tifinagh OCR Dataset (39k Synthetic Samples)

📌 Description

This dataset is a comprehensive collection of 39,104 synthetic images designed for training and evaluating OCR and vision-language models for the Tifinagh script. It features a wide variety of fonts, background colors, and text styles in a rectangular format.

📊 Dataset Composition

39,1K images in PNG format.
Diversity: Includes multiple background/text color combinations and various font styles (from: IRCAM, Tawalt, Madghis Madi).
Format: Optimized rectangular images (Dynamic scaling with centered text).

🧩 Features

Each sample in the metadata.jsonl contains:

file_name: Path to the image file.
text: Ground-truth Tifinagh transcription.
background_color: Color of the image background.
text_color: Color of the rendered text.
pic_ID: Unique image identifier.

The metadata has been shuffled randomly to allow a diverse preview of different fonts and styles in the Hugging Face dataset explorer.

🎯 Use Cases

Tifinagh OCR training and evaluation.
Document understanding for Amazigh languages.
Vision-language model validation.
Robustness testing across different typographic styles.

🛠 Usage

from datasets import load_dataset

dataset = load_dataset("Tamazight/Tifinagh-OCR-39K")

sample = dataset["train"][0]
print(sample["text"])
sample["image"].show()

📜 Metadata & Credits

Creation Year: 2024
License: apache-2.0
Creator: Aksel Tinfat for @Tamazight project on Hugging Face.