Darmm/darmm-ocr-kazakh-cyrillic

Darmm Kazakh Cyrillic OCR Dataset

Synthetic OCR dataset for Kazakh Cyrillic text. 200,000 rendered (image, label) pairs — word-level and line-level crops — generated from Kazakh Wikipedia text.

The first large-scale open OCR training dataset specifically targeting Kazakh Cyrillic script.

Dataset Summary

• 200,000 images total — 100k word-level + 100k line-level
• Train / Validation split — 190,000 / 10,000
• Script — Kazakh Cyrillic (includes Kazakh-specific characters: Ә, Ғ, Қ, Ң, Ө, Ұ, Ү, Һ, І)
• Text source — Kazakh Wikipedia (wikimedia/wikipedia, 20231101.kk)
• License — Apache 2.0

Data Fields

• image — PNG image of rendered text
• label — Ground truth string
• level — "word" or "line"

Generation

Images are synthetically rendered using Pillow with the following variation:

• Fonts — DejaVu Sans/Serif, FreeSans/Serif, Liberation Sans/Serif/Mono (16 variants)
• Font sizes — 16–48px (word), 14–36px (line)
• Backgrounds — white and light off-white variants
• Text colors — near-black variants
• Augmentation — Gaussian noise, slight blur, brightness/contrast jitter, small rotation (±2°)

Generation script: generate_kazakh_ocr.py

Motivation

Existing OCR systems regularly confuse Kazakh Cyrillic with Russian — the two scripts share most characters but Kazakh adds 9 additional letters (Ә, Ғ, Қ, Ң, Ө, Ұ, Ү, Һ, І) that standard Russian-trained OCR models handle poorly. No large-scale open training dataset existed for Kazakh OCR before this release.

Usage

from datasets import load_dataset

ds = load_dataset("Darmm/darmm-ocr-kazakh-cyrillic")

# Word-level only
word_ds = ds.filter(lambda x: x["level"] == "word")

# Line-level only
line_ds = ds.filter(lambda x: x["level"] == "line")

Intended Use

• Fine-tuning OCR models (TrOCR, PaddleOCR, EasyOCR) for Kazakh Cyrillic
• Benchmarking OCR systems on Kazakh text
• Pre-training visual text encoders for Kazakh

Limitations

• Synthetic only — does not cover handwritten text, scanned documents, or scene text
• Printed text on clean backgrounds — real-world documents may have more complex layouts
• Font diversity limited to system fonts available at generation time

Related Work

• KOHTD — Kazakh offline handwritten text dataset (3k exam papers, Cyrillic)
• henrygagnier/kazakh-ocr — Synthetic benchmark for Arabic/Cyrillic/Latin Kazakh OCR evaluation

Citation

@dataset{darmm_ocr_kazakh_cyrillic_2026,
  author    = {Zhumabek, Rakhat},
  title     = {Darmm Kazakh Cyrillic OCR Dataset},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/Darmm/darmm-ocr-kazakh-cyrillic}
}

Maintainer

R3iwan — part of the Darmm project.