DanishMahdi/snd_ocr_dataset

Sindhi OCR Dataset

Dataset Details

Dataset Description

This dataset is designed for Optical Character Recognition (OCR) research and development for the Sindhi language (Arabic script). It contains synthetically generated text-image pairs where each image consists of black Sindhi text rendered on a white background. The corresponding ground-truth text is provided for OCR model training and evaluation. It includes single words as well as short phrases formed by combining two or three dictionary words to increase lexical diversity and improve model generalization.

language:
  - sd

task_categories:
  - optical-character-recognition

task_ids:
  - image-to-text

splits:
  train: 256000
  validation: 8000
  test: 14000

total_samples: 278000

Data Generation

text_source:

• "Sindhi dictionary words"
• "Two-word combinations"
• "Three-word combinations"

image_format: background: "white" text_color: "black" image_type: "synthetically generated text images"

ground_truth:

• "Unicode Sindhi text"

Fonts Used

fonts:
  - "Lateef-Regular.ttf"
  - "Lateef-Bold.ttf"
  - "Lateef-Light.ttf"
  - "Lateef-Medium.ttf"
  - "Lateef-SemiBold.ttf"
  - "Lateef-ExtraBold.ttf"
  - "Lateef-ExtraLight.ttf"
  - "MB-Lateefi-SKv2.0.ttf"
  - "MBNargisNew3.2.ttf"

Dataset Structure

features:
  image:
    type: image
    description: "Black Sindhi text rendered on a white background"

  text:
    type: string
    description: "Ground-truth Sindhi text corresponding to the image"

Example:

image: "images/sample_00001.png"
text: "سنڌي ٻولي"

Intended Uses

suitable_for:

• Sindhi OCR model training
• Handwritten-to-printed OCR transfer research
• Vision-language model fine-tuning
• Text recognition benchmarking
• Low-resource language OCR research

not_recommended_for:

• Real-world document OCR evaluation without additional testing
• Historical document recognition
• Handwritten Sindhi OCR

Limitations

limitations:

• The dataset is synthetically generated and may not fully represent real-world scanned documents.
• Only black text on white backgrounds is included.
• The dataset does not contain handwritten text.
• Document layouts, tables, stamps, signatures, and complex formatting are not represented.

Citation

If you use this dataset in your research, please cite it as:

@dataset{mahdi2026sindhiocr,
  author       = {Danish Mahdi},
  title        = {Sindhi OCR Dataset},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/DanishMahdi/snd_ocr_dataset/}
}