riotu-lab/SAND-Extended

SAND: A Large-Scale Synthetic Arabic OCR Dataset

Overview

SAND (Synthetic Arabic OCR Dataset) is a large-scale, synthetically generated dataset designed for training and evaluating Optical Character Recognition (OCR) models for Arabic text. This dataset addresses the critical need for comprehensive Arabic text recognition resources by providing controlled, diverse, and scalable training data that simulates real-world book layouts.

Key Features

• Massive Scale: 843,622 document images containing approximately 690 million words
• Extensive Typographic Diversity: Ten distinct Arabic fonts covering a wide range of styles
• Structured Formatting: Designed to mimic real-world book layouts with consistent typography
• Clean Data: Synthetically generated with no scanning artifacts, blur, or distortions
• Content Diversity: Text spans multiple domains including culture, literature, Shariah, social topics, and more

Dataset Structure

The dataset is divided into ten splits based on font name:

• Amiri: Classical Naskh typeface inspired by early 20th century typography
• Sakkal Majalla: Widely used font in contemporary Arabic publishing
• Arial: Modern sans-serif font common in digital publications
• Calibri: Microsoft's default font representing contemporary digital typography
• Scheherazade New: Traditional-style font based on classical manuscript styles
• Jozoor Font: Decorative Arabic font with more stylized character forms
• Lateef: Extended Arabic script font designed for readability at small sizes
• Noto Naskh Arabic UI: Part of Google's Noto family, designed for user interfaces
• Thabit: Monospaced Arabic font for technical documentation
• Al Jazeera Arabic Regular: Based on the typography used by Al Jazeera media

📋 Sample Images

Each split contains data specific to a single font with the following attributes:

• image_name: Unique identifier for each image
• chunk: The text content associated with the image
• font_name: The font used in text rendering
• image_base64: Base64-encoded image representation

Content Distribution

Category	Number of Articles
Culture	13,253
Fatawa & Counsels	8,096
Literature & Language	11,581
Bibliography	26,393
Publications & Competitions	1,123
Shariah	46,665
Social	8,827
Translations	443
Muslim's News	16,725
Total Articles	133,105

Font Specifications

Font	Words Per Page	Font Size	Characteristics
Sakkal Majalla	50–300	14 pt	Contemporary publishing style
Arial	50–500	12 pt	Modern sans-serif
Calibri	50–500	12 pt	Contemporary digital document
Amiri	50–300	12 pt	Classical Naskh typeface
Scheherazade New	50–250	12 pt	Traditional manuscript style
Noto Naskh Arabic UI	50–400	12 pt	Clear UI rendering
Lateef	50–350	14 pt	Optimized for small sizes
Al-Jazeera-Arabic	50–250	12 pt	Media/journalistic style
Thabit	50–240	12 pt	Monospaced technical font
Jozoor Font	50–200	12 pt	Decorative with stylization

Page Layout

Specification	Measurement
Page Size	A4 (8.27 × 11.69 in)
Left Margin	0.9 in
Right Margin	0.9 in
Top Margin	1.0 in
Bottom Margin	1.0 in
Gutter Margin	0.2 in
Resolution	300 DPI
Color Mode	Grayscale
Page Direction	Right-to-Left
Text Alignment	Right
Line Spacing	1.15

Usage Example

from datasets import load_dataset
import base64
from io import BytesIO
from PIL import Image
import matplotlib.pyplot as plt

# Load dataset with streaming enabled
ds = load_dataset("riotu-lab/SAND-Extended", streaming=True)
print(ds)

# Iterate over a specific font dataset (e.g., Amiri)
for sample in ds["Amiri"]:
    image_name = sample["image_name"]
    chunk = sample["chunk"]  # Arabic text transcription
    font_name = sample["font_name"]
    
    # Decode Base64 image
    image_data = base64.b64decode(sample["image_base64"])
    image = Image.open(BytesIO(image_data))

    # Display the image
    plt.figure(figsize=(10, 10))
    plt.imshow(image)
    plt.axis('off')
    plt.title(f"Font: {font_name}")
    plt.show()

    # Print the details
    print(f"Image Name: {image_name}")
    print(f"Font Name: {font_name}")
    print(f"Text Chunk: {chunk}")
    
    # Break after one sample for testing
    break

Working with Multiple Fonts

To train or evaluate models across different font styles:

from datasets import load_dataset
import random

# Load the dataset
ds = load_dataset("riotu-lab/SAND")

# Select a balanced sample from multiple fonts
fonts_to_use = ["Amiri", "Arial", "Scheherazade_New", "Thabit", "Noto_Naskh_Arabic_UI"]
samples_per_font = 1000
combined_samples = []

for font in fonts_to_use:
    # Get random samples from this font
    font_samples = ds[font].shuffle(seed=42).select(range(samples_per_font))
    combined_samples.extend([(sample, font) for sample in font_samples])

# Shuffle the combined samples
random.shuffle(combined_samples)

# Now you can use these samples for training or evaluation
for sample, font in combined_samples[:5]:  # Just show first 5 as example
    print(f"Font: {font}, Image: {sample['image_name']}")

Applications

SAND is designed to support various Arabic text recognition tasks:

• Training and evaluating OCR models for Arabic text
• Developing vision-language models for document understanding
• Fine-tuning existing OCR models for better Arabic script recognition
• Benchmarking OCR performance across different fonts and layouts
• Research in Arabic natural language processing and computer vision
• Developing font-adaptive OCR systems that generalize across typographic styles

Citation

If you use SAND in your research, please cite:

@misc{sand2025,
  title={SAND: A Large-Scale Synthetic Arabic OCR Dataset for Vision-Language Models},
  author={RIOTU Lab},
  year={2025},
  howpublished={\url{https://huggingface.co/datasets/riotu-lab/SAND}}
}

Acknowledgments

The authors thank Prince Sultan University for their support in developing this dataset.