Urdu OCR Dataset with 1.5 Million Synthetic Samples

A large-scale synthetic dataset for Urdu Optical Character Recognition (OCR) created by PuristanLabs1. It contains 1.5 million samples, including 499,845 rendered with authentic Jameel Noori Nastaliq ligatures and 1,000,160 in standard Urdu Naskh fonts. The dataset was last updated on February 1,我们发现了一个问题。

Use Cases

• Train OCR models for Urdu text recognition based on the described Nastaliq and Naskh samples.
• Benchmark OCR performance on the Nastaliq script based on the authentic ligatures mentioned.
• Develop synthetic data pipelines for text generation based on the custom rendering pipeline described.
• Fine-tune multilingual OCR systems to include Urdu based on the large-scale sample count.

Strengths

• Contains 1.5 million total samples, providing a large-scale resource.
• Includes 499,845 samples with authentic Jameel Noori Nastaliq ligatures, addressing a specific script style.
• Provides a robust baseline with 1,000,160 samples in standard Urdu Naskh fonts.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• The dataset is synthetic, which may not fully capture the noise and variation of real-world documents.

Provenance

Source

PuristanLabs1 on Hugging Face.

Collection Method

Synthetically generated using a custom Chromium-based rendering pipeline.

Time Range

null

Freshness

Last updated 2026-02-01 08:45:40; freshness should be verified.

Geography

null