README
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 or the accompanying TurkicOCR-SVTRv2-B recognizer in your research, please cite:
@inproceedings{issayev2026turkicocr,
title={TurkicOCR-SVTRv2-B: Lightweight Line-Grounded Recognizer for Kazakh and Kyrgyz Optical Character Recognition},
author={Issayev, Alen and Zhalgas, Aidana},
booktitle={Analysis of Images, Social Networks and Texts (AIST 2026)},
series={Lecture Notes in Computer Science (LNCS)},
publisher={Springer},
year={2026},
doi={10.1007/978-3-031-XXXXX-X_XX}
}
@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}
}
Similar repos
README
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 or the accompanying TurkicOCR-SVTRv2-B recognizer in your research, please cite:
@inproceedings{issayev2026turkicocr,
title={TurkicOCR-SVTRv2-B: Lightweight Line-Grounded Recognizer for Kazakh and Kyrgyz Optical Character Recognition},
author={Issayev, Alen and Zhalgas, Aidana},
booktitle={Analysis of Images, Social Networks and Texts (AIST 2026)},
series={Lecture Notes in Computer Science (LNCS)},
publisher={Springer},
year={2026},
doi={10.1007/978-3-031-XXXXX-X_XX}
}
@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}
}