alenisaw/turkicocr-cyrillic

TurkicOCR Synthetic Cyrillic Dataset

A large-scale synthetic dataset for document AI research in underrepresented Turkic languages — Kazakh and Kyrgyz. Built to cover the full document understanding pipeline: text detection, recognition (OCR), layout analysis, and visual document understanding (VDU).

Pages span 29 authentic document archetypes across administrative, educational, and commercial domains, rendered with 7 procedural degradation profiles that simulate real-world capture conditions — from clean office prints to aged paper, phone photographs, and official ink stamps.

Three nested configs (tiny / medium / large) enable progressive training scale without re-downloading data.

from datasets import load_dataset
ds = load_dataset("alenisaw/turkicocr-cyrillic", name="large")

Configs

Config	Total	Train	Validation	Test
tiny	25,000	22,500	1,250	1,250
medium	50,000	45,000	2,500	2,500
large	100,000	90,000	5,000	5,000

tiny ⊂ medium ⊂ large — deterministic nested views of the same generation. Images are stored as JPEG inside packed TAR shards; parquet indexes reference each page by page_id and tar_path.

Document Layouts

29 layouts across 5 categories:

Category	Layouts
Administrative & Official	Official letters, memos, meeting minutes, official statements, archival notifications, certificates
Forms & Registries	Application forms, simple forms, registry extracts
Books & Prose	Single/two-column book pages, dictionary entries, glossaries, indexes, academic abstracts, bulletins, historical newspapers
Educational & Specialized	Syllabi, lecture notes, exam sheets, exam registers, worksheets
Tables & Transactional	Invoices, receipts, catalog entries, attendance/schedule/simple/wide-schedule tables, inventory sheets

Degradation Profiles

7 procedurally generated visual effect profiles:

Profile	Simulates
clean	No degradation
low_dpi_scan	Low-resolution scan artifacts
office_scan	Office scanner noise and banding
official_stamped	Round/rectangular ink stamps and handwritten signatures
old_paper	Aging, yellowing, water stains, blotches
phone_photo	Perspective distortion, lens blur, camera projection
photocopy	Repeated photocopy erosion and thresholding

Intended Use

For training and evaluating OCR, document layout analysis, and VDU models (LayoutLM, Donut, Pix2Struct, ColPali). The dataset is synthetic — validate on real-world documents before deployment.

Limitations

• Synthetic content: Text is procedurally generated from corpus sources. Semantic coherence between entities (e.g. name–address binding on forms) is not guaranteed. Optimized for visual/geometric recognition, not semantic NLP tasks.
• Domain gap: Real-world generalization should be verified on actual scanned or photographed documents.

Acknowledgements

The author would like to thank the Research and Innovation Center "CyberTech" at Astana IT University for their support and resources during the creation of this dataset.

Citation

If you use this dataset, please cite it as:

@misc{issayev_2026_turkicocr_cyrillic,
  author       = {Issayev, Alen},
  title        = {TurkicOCR-Cyrillic},
  year         = {2026},
  publisher    = {Hugging Face},
  doi          = {10.57967/hf/9255},
  url          = {https://huggingface.co/datasets/alenisaw/turkicocr-cyrillic},
  note         = {Synthetic Cyrillic OCR and document-understanding dataset}
}