600k_KS_OCR_Word_Segmented_Dataset

A large-scale synthetic OCR dataset for Kashmiri script containing approximately 602,000 word-level segmented images designed for training and evaluating OCR models, particularly CRNN and TrOCR architectures.

Dataset Overview

Property	Value
Total Samples	~602,000 words
Archive Files	10 ZIP files
Image Dimensions	256×64 pixels (CRNN Standard) / TrOCR
Text Direction	RTL (Right-to-Left)
Script	Kashmiri (Perso-Arabic)
Formats	CRNN, TrOCR, CSV, JSONL

Image	Transcript
	پہلگام
	چہِ
	فُٹ
	بال
	فیضہ
	خانَن
	کوٚر
	پہلگام
	سپورٹس
	گراؤنڈَس
	شوٗٹِنٛگ
	مُقابلَس
	منٛز
	سۄنہٕ
	سُنٛد
	تمغہٕ،
	رِیٲستی

Dataset Structure

The dataset is distributed across 10 ZIP archives:

600k_KS_OCR_Word_Segmented_Dataset/
├── P1_OCR_dataset_50000_2025-12-26T11-26-04.zip    (~904 MB)
├── P2_OCR_dataset_53815_2025-12-26T12-17-02.zip    (~953 MB)
├── P3_OCR_dataset_68741_2025-12-26T12-32-17.zip    (~1.2 GB)
├── P4_OCR_dataset_69886_2025-12-26T12-43-47.zip    (~1.2 GB)
├── P5_OCR_dataset_69637_2025-12-26T13-01-04.zip    (~1.2 GB)
├── P6_OCR_dataset_69506_2025-12-26T13-14-06.zip    (~1.2 GB)
├── P7_OCR_dataset_58228_2025-12-26T14-05-51.zip    (~1.1 GB)
├── P8_OCR_dataset_35720_2025-12-26T15-31-59.zip    (~620 MB)
├── P9_OCR_dataset_86635_2025-12-26T16-14-27.zip    (~1.5 GB)
└── P10_OCR_dataset_41401_2025-12-26T16-25-16.zip   (~732 MB)

Contents of Each ZIP Archive

Each archive contains:

P*_OCR_dataset_*/
├── images/           # Folder containing word-segmented images
├── data.csv          # CSV format labels (filename, text)
├── data.jsonl        # JSONL format for TrOCR training
├── labels.txt        # CRNN format labels (filename \t text)
└── metadata.json     # Generation configuration and statistics

Generation Configuration

Image Settings

Parameter	Value
Size Preset	CRNN Standard
Dimensions	256×64 pixels
Background	#FFFFFF (White)
Background Style	Mixed (clean_white + textures)
Text Color	#000000 (Black)
Direction	RTL

Fonts Used

Three traditional Kashmiri/Nastaliq fonts were used for text rendering:

• Afan_Koshur_Naksh - Native Kashmiri Naskh style
• Nastaleeq - Classical Nastaliq calligraphy
• Nakash (Narqalam) - Traditional handwritten style

Data Augmentation

Augmentation was enabled with 60% augmentation percentage, applying various transformations to improve model robustness:

Augmentation Type
Rotation/Tilt
Skew/Perspective
Gaussian Blur
Motion Blur
Gaussian Noise
Salt & Pepper
Brightness Variation
Contrast Variation
JPEG Artifacts
Resolution Loss
Paper Texture
Shadow/Lighting
Ink Bleed

Note: Different ZIP files may have varying augmentation levels applied during generation.

Background Textures

Mixed background mode with diverse paper textures:

• clean_white - Pure white background
• aged_paper, old_book, parchment - Aged document styles
• notebook, book_page, newspaper - Document styles
• ivory, cream, recycled - Paper tones
• coffee_stain, weathered - Distressed effects
• p1-p14 background - Custom pattern backgrounds

Sample Metadata

Each ZIP contains a metadata.json with generation details:

{
  "generated_at": "2025-12-26T16:24:34.803Z",
  "config": {
    "image_size": "256x64",
    "background": "#FFFFFF",
    "background_style": "clean_white",
    "background_mode": "mix",
    "text_color": "#000000",
    "direction": "rtl",
    "augmentation_enabled": true,
    "augmentation_percentage": 60,
    "fonts_used": [
      "Afan_Koshur_Naksh",
      "Nastaleeq",
      "Nakash (Narqalam)"
    ],
    "output_formats": ["crnn", "trocr", "csv", "jsonl"]
  },
  "samples": ,
  "clean_samples": ,
  "augmented_samples": ,
  "font_usage": {
    "Nakash (Narqalam)": ,
    "Afan_Koshur_Naksh": ,
    "Nastaleeq": 
  },
  "background_usage": {
    "clean_white": ,
    "p3 background": ,
    "p2 background": ,
    ...
  }
}

Data Formats

labels.txt (CRNN Format)

image_001.png	کٲشُر
image_002.png	زَبان

data.jsonl (TrOCR Format)

{"file_name": "image_001.png", "text": "کٲشُر"}
{"file_name": "image_002.png", "text": "زَبان"}

data.csv

filename,text
image_001.png,کٲشُر
image_002.png,زَبان

Performance Notes

Generated using GPU-accelerated parallel processing:

• Parallel Batches: 4 cores
• Performance Boost: 3-10x faster generation
• GPU Augmentation: Enabled

Usage

Loading with Hugging Face Datasets

from datasets import load_dataset

dataset = load_dataset("Omarrran/600k_KS_OCR_Word_Segmented_Dataset")

Manual Loading

import json
import pandas as pd
from pathlib import Path

# Load labels
labels_df = pd.read_csv("data.csv")

# Load metadata
with open("metadata.json", "r") as f:
    metadata = json.load(f)

# Load images
from PIL import Image
img = Image.open("images/image_001.png")

Intended Use

• Training OCR models for Kashmiri script recognition
• Fine-tuning TrOCR, CRNN, and attention-based sequence models
• Benchmarking low-resource language OCR systems
• Research in Perso-Arabic script recognition

Language Information

Kashmiri (کٲشُر) is an endangered Dardic language spoken primarily in the Kashmir Valley. It uses a modified Perso-Arabic script with additional characters for unique phonemes. This dataset aims to advance OCR technology for this underrepresented language.

Citation

If you use this dataset, please cite:

@dataset{ks_ocr_600k_2025,
  title={600k Kashmiri OCR Word Segmented Dataset},
  author={Haq Nawaz Malik},
  year={2025},
  publisher={Hugging Face},
  published url={https://huggingface.co/datasets/Omarrran/600k_KS_OCR_Word_Segmented_Dataset}}
}

License

[ CC-BY-4.0]

Dataset & License Policy (v1.0)

© Copyright Notice

All datasets, derived datasets, annotations, and language resources related to the Kashmiri AI initiative are © the Dataset Creator / Project Maintainer unless otherwise stated.

---

PART A — DATASET LICENSE

(Controlled Research Access License )

1. Definitions

• “Datasets” refers to all text, audio, OCR, parallel corpora, lexicons, annotations, and derived linguistic resources.
• “Derived Data” means any data generated directly or indirectly using the original datasets.
• “User” means any individual, institution, or organization accessing the datasets.

---

2. Ownership

• Full ownership of the datasets remains with the Dataset Creator.
• Access to datasets does not transfer ownership, copyright, or exclusive rights.

---

3. Permitted Uses

Users may use the datasets for:

1. Non-commercial research and experimentation
2. Academic publications and benchmarking
3. Training AI models for Kashmiri language research
4. Open scientific collaboration aligned with language preservation
5. Internal evaluation and analysis

All permitted uses require proper attribution.

---

4. Prohibited Uses

Users must NOT:

1. Redistribute, mirror, or re-host datasets (publicly or privately)
2. Use datasets for commercial products, services, APIs, or SaaS
3. Train proprietary or closed-source models without permission
4. Attempt dataset reconstruction via model outputs
5. Combine datasets into other datasets for redistribution
6. Claim ownership or exclusive rights over the data
7. Use datasets in surveillance, profiling, or harmful applications

Violations result in immediate revocation of access.

---

5. Model Training Restrictions

• Models trained using these datasets:

  • Must clearly acknowledge dataset usage
  • Must not enable extraction or regeneration of original data
  • May not be commercialized without a separate agreement

• Weights may be:

  • Open
  • Gated
  • Restricted at the discretion of the maintainer.

---

6. Attribution Requirements

All uses must include:

“If you use this corpus in your research or applications, please cite as per above Citaion section:

”

Required in:

• Research papers
• Model cards
• GitHub repositories
• Public demos or reports

Failure to attribute constitutes a license violation.

---

7. Access Control

Dataset access may be:

• Revoked at any time
• Limited by scope or duration
• Subject to additional conditions

No guarantee of permanent access is implied.

---

8. Commercial Licensing

Any commercial use requires:

• Explicit written permission
• A separate licensing agreement
• Possible royalties or revenue-sharing

Commercial inquiries must be made before use.

---

9. Disclaimer

Datasets are provided “AS IS”, without warranty of any kind. The maintainer is not liable for downstream use, misuse, or consequences.

---

---

10. Termination

This license is automatically terminated if:

• Any term is violated
• Attribution is removed
• Datasets are misused

Upon termination, all copies must be deleted.

---

11. Acceptance

By accessing datasets or using tools, you agree to all terms in this policy.

---

Contact

For questions or issues, please open an issue on the repository or contact the dataset creator.

---

This dataset was generated as part of ongoing efforts to develop NLP and OCR resources for the Kashmiri language.