PiotrSty/impact-psnc-polish-ocr

IMPACT-PSNC Polish OCR Diverse Subset

Compact, provenance-preserving subset of the Polish IMPACT ground truth released by the Poznan Supercomputing and Networking Center (PSNC). It is intended for OCR experiments on diverse historical Polish printed material.

This subset contains:

• 89 full-page images from 30 source collections;
• 599 text-region crops derived from PAGE XML polygons;
• the 89 corresponding original PAGE XML files;
• page and region transcriptions;
• document-level train, validation, and test splits;
• source URLs and SHA-256 checksums.

The JPEG files preserve the source pixel dimensions and were derived from the original TIFF files at quality 95 with no chroma subsampling. PAGE coordinates therefore remain applicable to the page images.

Configurations

from datasets import load_dataset

pages = load_dataset("PiotrSty/impact-psnc-polish-ocr", "pages")
regions = load_dataset("PiotrSty/impact-psnc-polish-ocr", "regions")

The pages configuration includes page images, full transcriptions, PAGE XML paths, dimensions, source identifiers, source URLs, and checksums. The regions configuration contains recognition-ready text crops, transcriptions, region types, polygons, bounding boxes, and parent page identifiers.

Selection and splits

Three annotated pages were selected at approximately the beginning, middle, and end of each source collection. One requested page returned HTTP 401, resulting in 89 rather than 90 pages.

Splits are assigned by whole source collection, so neighboring pages from one work do not cross split boundaries:

Split	Pages	Regions	Collections
train	65	416	22
validation	15	116	5
test	9	67	3

Zbior_rytmow_FT was excluded because its current dLibra image identifiers could not be mapped deterministically from the PAGE XML filenames. Oekonomika_FT/00436449.xml was excluded because its image endpoint returned HTTP 401. These exclusions are recorded in stats.json.

Source and provenance

• PAGE XML source: impactcentre/groundtruth-pol
• Pinned source commit: 0a5a88863f104d8ab26b22b7c0a32c58dd3c5f15
• Image and metadata source: IMPACT dLibra
• Source institution: Poznan Supercomputing and Networking Center (PSNC)

The original repository states that these Polish Digital Library ground-truth files are released under CC BY. The dLibra records used here independently label both the attachment and the object licence as CC BY.

See source_manifest.json, NOTICE.md, and checksums.sha256 for exact provenance and integrity information.

License and attribution

The source dataset is distributed under Creative Commons Attribution 3.0 Unported. Cite and attribute the IMPACT Centre of Competence and Poznan Supercomputing and Networking Center (PSNC), and retain a link to the license. The JPEG images and region crops in this repository are format-converted derivatives of the licensed source images and remain under CC BY 3.0.

Suggested attribution:

IMPACT Polish ground truth, Poznan Supercomputing and Networking Center (PSNC), via the IMPACT Centre of Competence, CC BY 3.0.

Limitations

• This is historical printed Polish OCR data, not handwriting data.
• Transcriptions preserve historical spelling, long-s and other historical glyphs, and occasional replacement characters present in the source PAGE XML.
• Text regions are rectangular crops around PAGE polygons and may include neighboring content on complex layouts.
• The subset is designed for diversity and lightweight experimentation, not to reproduce the full source distribution.
• Do not treat the provided test split as a universal benchmark; it contains only three held-out source collections.

Reproducibility

The selection, conversion, crop generation, source exclusions, and validation results are recorded in scripts/, stats.json, source_manifest.json, and validation_report.json.