You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

Arabic Salary Report OCR Dataset β€” Mixed Numeric Formats

πŸ“„ Overview

The Arabic Salary Report OCR Dataset is a synthetic dataset of 30,000 images designed for training and evaluating OCR systems on Arabic text containing numeric data.
It incorporates variations in numeric representation, including both Arabic-Indic numerals (Ω Ω‘Ω’Ω£Ω€Ω₯Ω¦Ω§Ω¨Ω©) and Western numerals (0–9), embedded in realistic salary report layouts.

This dataset is ideal for:

  • Fine-tuning OCR models to recognize Arabic salary reports.
  • Handling mixed-language numeric formats.
  • Benchmarking Arabic financial document parsing.

πŸ“¦ Dataset Composition

The dataset contains 30,000 images split into two main structural formats:

Format Type Quantity Description
Table format 15,000 Salary figures embedded inside structured tables.
Paragraph format 15,000 Salary figures integrated into continuous Arabic text paragraphs.

Each format has an even split of numeric styles:

  • 50% Arabic-Indic numerals only.
  • 50% Mixed numerals (combination of Arabic-Indic and Western).

All text content is entirely in Arabic, except for the Western numerals in the mixed format.


πŸ›  Data Generation & Purpose

The dataset was synthetically generated to simulate realistic salary reports, ensuring:

  • Variation in font styles, sizes, and layouts.
  • Presence of both structured (tables) and unstructured (paragraphs) salary data.
  • Representation of both numeric systems to improve OCR model robustness.

πŸ” Example Use Cases

  • Training OCR models to handle Arabic text with mixed numerals.
  • Fine-tuning language models for salary extraction from financial documents.
  • Benchmarking document understanding systems for Arabic financial reports.

πŸ“₯ Usage

To load the dataset in Python with Hugging Face datasets:

from datasets import load_dataset

dataset = load_dataset("moekh/new-digit-ocr-dataset")
print(dataset)
Downloads last month
24