AlexandreSheva/rukopys-curated-mvp-v2

RUKOPYS Curated MVP: Ukrainian Handwriting Recognition Dataset

RUKOPYS Curated MVP is a cleaned, task-ready derivative of UkrainianCatholicUniversity/rukopys for Ukrainian handwritten document AI. It turns the raw RUKOPYS release into reproducible artifacts for page-level vision-language fine-tuning, crop-level transcription, and layout detection.

This dataset is designed for practical HTR work: train a model, inspect the normalized records, evaluate layout/text extraction, and export Kaggle-style predictions from the same data contract.

What This Dataset Enables

• Full-page image-to-JSON training for models that return text regions with bounding boxes.
• Crop-level handwritten text transcription fine-tuning.
• YOLO-format layout detection experiments.
• Consistent metadata for evaluation, debugging, and model-card traceability.
• Reproducible packaging for Hugging Face Hub uploads and local downloads.

Dataset Contents

• metadata.jsonl: normalized page records with image metadata and region annotations.
• regions.jsonl: one row per annotated region for analysis and filtering.
• vlm_sft.jsonl: crop-level supervised examples for handwritten text transcription.
• page_sft.jsonl: full-page image-to-structured-JSON supervised examples.
• page_text_sft.jsonl: compact full-page image-to-reading-order text-line examples.
• yolo/: YOLO-format layout detection dataset and split metadata.
• crops/: region crops for transcription fine-tuning, when crop export is enabled.
• _pack_manifest.json: tar-shard manifest when the dataset is uploaded in packed form.

Use rukopys download-curated to download the dataset and automatically unpack tar shards.

Dataset Size

• Pages: 2715
• Regions: 39420
• Crop-level SFT examples: 38832
• Page-level SFT examples: 2330
• Page-text SFT examples: 2309

Source Distribution

• archive: 178
• dictation: 509
• school: 1782
• university: 246

Region Types

• annotation: 1145
• formula: 6174
• graph: 61
• handwritten: 31193
• image: 130
• printed: 477
• table: 240

Quality Signals

Each example receives a quality weight derived from annotation provenance. The weights are intended for filtering or weighted sampling during training:

• annotator: 1.0
• volunteer: 0.75
• auto: 0.35

These weights do not replace evaluation; they provide a simple way to reduce the impact of noisier silver or automatic annotations.

Recommended Uses

• Fine-tune VLMs for Ukrainian handwritten page parsing.
• Train layout detectors for handwritten and printed regions.
• Benchmark OCR/HTR post-processing on structured document outputs.
• Demonstrate an end-to-end document AI pipeline from raw data to model submission artifacts.

Limitations

• This is a curated derivative, so it inherits the source dataset's annotation assumptions and gaps.
• Automatic or volunteer-derived annotations can be noisy; use quality_weight when training.
• The dataset is focused on Ukrainian handwriting and should not be treated as language-general OCR.
• Page scans vary in quality, layout, and handwriting style, so validation on target documents is required.

License

This dataset inherits the source dataset license: CC BY-NC-SA 4.0. Check the license terms before commercial or redistribution use.

Source Dataset

This dataset is derived from UkrainianCatholicUniversity/rukopys.

@dataset{rukopys_2026,
  title        = {RUKOPYS}: Ukrainian Handwritten Text Recognition Dataset,
  author       = {Dmytro Voitekh and Volodymyr Zmiivskyyi and Oleksii Molchanovskyi},
  organization = {Ukrainian Catholic University},
  year         = {2026},
  license      = {CC BY-NC-SA 4.0},
  url          = {https://huggingface.co/datasets/UkrainianCatholicUniversity/rukopys},
  note         = {First large-scale Ukrainian HTR dataset; from 1920s archival documents to 2025 school homework and exams}
}