---
dataset_info:
  features:
  - name: id
    dtype: int64
  - name: image
    dtype: image
  - name: text
    dtype: string
  - name: bbox
    dtype: string
  - name: lines
    dtype: string
  - name: tokens
    dtype: string
  - name: theme
    dtype: string
  - name: font
    dtype: string
  - name: font_size
    dtype: int64
  - name: font_weight
    dtype: int64
  - name: font_color
    dtype: string
  - name: line_height
    dtype: int64
  - name: bg_color
    dtype: string
  - name: justify
    dtype: string
  - name: justify_last_line
    dtype: string
  splits:
  - name: test
    num_bytes: 160016504.0
    num_examples: 10000
  download_size: 142306866
  dataset_size: 160016504.0
configs:
- config_name: default
  data_files:
  - split: test
    path: data/test-*
license: mit
task_categories:
- text-generation
language:
- ps
- ar
- en
- fa
tags:
- ocr
- nlp
- cv
size_categories:
- 10K<n<100K
---

# PsOCR - Pashto OCR Dataset

<p align="center">
    <img src="logo.png" width="200"/>
<p>
<p align="center">
        🌐 <a href="https://zirak.ai/"><b>Zirak.ai</b></a>
          &nbsp&nbsp | &nbsp&nbsp🤗 <a href="https://huggingface.co/datasets/zirak-ai/Pashto-OCR"><b>HuggingFace</b></a>
          &nbsp&nbsp | &nbsp&nbsp <a href="https://github.com/zirak-ai/PashtoOCR"><b>GitHub</b></a>
          &nbsp&nbsp | &nbsp&nbsp <a href="https://www.kaggle.com/datasets/drijaz/PashtoOCR"><b>Kaggle</b></a>
          &nbsp&nbsp | &nbsp&nbsp📑 <a href="https://doi.org/10.1016/j.asej.2026.104024"><b>Paper</b></a>
</p>

**[PsOCR: Benchmarking Large Multimodal Models for Optical Character Recognition in Low-resource Pashto Language](https://doi.org/10.1016/j.asej.2026.104024)**

The dataset is also available at: https://www.kaggle.com/datasets/drijaz/PashtoOCR

# Introduction
* PsOCR is a large-scale synthetic dataset for Optical Character Recognition in low-resource Pashto language. 
* This is the first publicly available comprehensive Pashto OCR dataset consisting of **One Million** synthetic images annotated at word, line, and document-level granularity, covering extensive variations including **1000** unique font families, diverse colors, image sizes, and text layouts.
* PsOCR includes the first publicly available OCR benchmark comprising **10,000** images, facilitating systematic evaluation and comparison of OCR systems for the low-resource Pashto.
* We conducted a pioneering evaluation and comparison of state-of-the-art LMMs on Pashto OCR, providing crucial insights into their zero-shot capabilities, strengths, and limitations for low-resource languages written in Perso-Arabic scripts.
* <span style="color:red"><b>ℹ️ On this repo, only the test set (benchmark) is available.  If you need the full dataset of 1M images, please contact us.</b></span>

<p align="center">
    <img src="fig1.jpg" width="50%"/><br>
    Performance Comparison of various LMMs on PsOCR Benchmark
<p>

# Granularity
The annotation information is provided at three levels of granularity: page-level, line-level, and token-level

<p align="center">
    <img src="fig2.jpg" width="100%"/>
<p>

# Font Variation
PsOCR features 1000 unique font families, a few of them are shown here.
<p align="center">
    <img src="fig4.jpg" width="100%"/>
<p>

# Citation

If you found our work useful, please feel free to cite it:

```bibtex
@article{Haq2026PsOCR,
  title   = {PsOCR: Benchmarking Large Multimodal Models for Optical Character Recognition in Low-Resource Pashto Language},
  journal = {Ain Shams Engineering Journal},
  volume  = {17},
  number  = {3},
  pages   = {104024},
  year    = {2026},
  issn    = {2090-4479},
  doi     = {10.1016/j.asej.2026.104024},
  url     = {https://www.sciencedirect.com/science/article/pii/S2090447926000511},
  author  = {Ijazul Haq and Yingjie Zhang and Muhammad Saqib}
}
```
# Contact
**Website:** https://zirak.ai/  
**Email Address:** [contact@zirak.ai](mailto:contact@zirak.ai), [mail@ijaz.me](mailto:mail@ijaz.me)