tahaListens/seraiki-handwritten-numerals

📊 Seraiki Handwritten Numbers (1–99)

Dataset Summary

This dataset contains handwritten Seraiki numbers from 1 to 99, written in words using the Perso-Arabic Seraiki script.
It was created to support Optical Character Recognition (OCR) research and to promote AI development for the underrepresented Seraiki language.

Unlike digit-only datasets (e.g., MNIST), this dataset includes full word forms of numbers, making it suitable for sequence-based recognition tasks (TrOCR, CRNN, VisionEncoderDecoder).

---

Dataset Details

• Languages: Seraiki (سرائیکی)
• Script: Perso-Arabic (same family as Urdu script)
• Labels: Word form of numbers 1–99
• Features:
  • image: handwritten number in Seraiki script
  • text: transcription in Seraiki

Example:

{
  "image": <PIL.Image.Image>,
  "text": "پنج"   // "five"
}
Splits
train: 80%

test: 20%

Supported Tasks
Handwritten Word Recognition (OCR)

Image-to-Text (TrOCR, Donut, Encoder–Decoder models)

Benchmarking on low-resource handwriting datasets

Evaluation Metrics:

Khan, Muhammad Ahmad (2023), “Saraiki Language Character Dataset”, Mendeley Data, V1, 

doi: 10.17632/tc9zv2wf2k.1



---
license: cc-by-4.0
task_categories:
- image-to-text
tags:
- saraiki
- seraiki
- pakistan
---