This is a large-scale, synthetically generated dataset designed for Urdu OCR (Optical Character Recognition), specifically focusing on the Nastaliq script. It contains 50,000 high-quality, full-page images of Urdu Shayari (Poetry) along with their exact ground-truth text transcriptions.

This dataset is intended to fine-tune Vision-Language Models (VLMs) like Qwen2-VL-2B, Qaari-Urdu, or TrOCR for extracting text from complex Urdu book pages.

Dataset Structure

Columns

image: A PIL Image object representing a full scanned book page.
text: A string containing the exact Urdu transcription of the page (including line breaks for couplets).

Features

Font: Jameel Noori Nastaleeq (a highly complex and elegant Nastaliq font).
Layout: Book-style pages with right-aligned text, proper margins, and distinct gaps between couplets (Ashaar).
Backgrounds (Mixed):
- 33% Clean white paper (modern print).
- 33% Light cream paper (classic book).
- 34% Old/dusty yellowed paper (scanned vintage look).
Page Numbers: Randomly generated page numbers at the bottom center of most pages.
Augmentations: Slight rotation, dust specks, and realistic ink bleed to simulate real-world scans.

How to Use

You can easily load this dataset using the datasets library. Here is a quick example:

from datasets import load_dataset

dataset = load_dataset("YOUR_HF_USERNAME/YOUR_DATASET_NAME")
sample = dataset['train'][0]

# Display image and text
display(sample['image'])
print(sample['text'])

Downloads last month: 157

Total file size:

6 GB