bicolino34/uk-wikisource-OCR-full

Ukrainian Wikisource Page Transcriptions

Page images from digitized books and periodicals paired with their human-made transcriptions from Ukrainian Wikisource. It can be used to train and evaluate OCR / document-understanding models, especially for Ukrainian texts in historical orthographies.

The dataset covers almost all pages that have both a scan and a transcription on Ukrainian Wikisource at the time of collection (2026.10.01), although few hundred pages were dropped.

Dataset structure

Each row is one page.

Column	Type	Description
Title	string	Page name from where the text is extracted in Ukrainian Wikisource.
Text	string	Raw wikitext of the page, exactly as written by Wikisource editors, including templates, tables and <ref> footnotes.
Image_Filename	image	Scan of the page, as stored in the Wikisource Сторінка: (Page) namespace.
Text_Clean	string	Automatically cleaned version of Text column (see below).

Content

• Language: primarily Ukrainian. The texts span many periods and various orthographies (e.g. early phonetical and etymological orthographies, Galician/Austrian-era spellings, Kulishivka, the 1928 Kharkiv orthography, Soviet and modern standards), so spelling varies from edition to edition.
• Other languages: editions often contain fragments in other languages (Russian, Polish, German, French, Latin, etc.), usually only short passages or only a few pages.
• Material: books, brochures, periodicals and other printed works, in a range of typefaces and scan qualities. Some pages contain tables, footnotes, headers or illustrations.
• Scripts: mostly Cyrillic, with some Latin.

The two text columns

Text

The original page source from Wikisource. It contains MediaWiki markup that mostly encodes layout and formatting rather than text, for example {{center|...}}, {{sp|...}}, page header/footer templates (in element at the start and at the end of page), and so on. Use it if you want the unmodified editor input or want to apply your own cleaning.

Text_Clean

A version with wikitemplates removed by a rule-based script. Best effort; it is not perfect. What it does:

• Removes templates (including nested ones and multi-line ones). For pure formatting wrappers (centering, letter-spacing, font size, language tags, etc.) the visible text is kept and only the wrapper is removed.
• Keeps footnotes: <ref>...</ref> and <ref ... /> elements are preserved because they are transcribed inline inside the paragraphs (templates inside them are cleaned as well). You can easily clean them depending on your needs.
• Keeps wikitables, with styling and cell attributes stripped; rowspan / colspan are preserved, since they affect the table structure. For more details on tables see https://w.wiki/eKsp
• Removes HTML comments, leftover <noinclude>-type tags and extra blank lines.

Known limitations: some templates may have been dropped even though their content was visible text, other templates may have been unwrapped incorrectly, and rare markup patterns may slip through. If you need maximum reliability for a specific use, compare against Text.

Intended uses

• Training and evaluating OCR / HTR and vision-language models on Ukrainian and historical-orthography text.

Limitations and biases

• Transcriptions are written and proofread by volunteers; quality and consistency vary between pages and editions. Original orthography is generally preserved, but editors sometimes make different choices.
• Not every transcription is guaranteed to match its image perfectly; some pages have only been checked by one person while others have been approved by at least two different people (split is around 50-50).
• The collection reflects what volunteers chose to digitize, so it is skewed toward certain periods, genres and regions.
• Some pages are almost blank (covers, separators, plates) and have short or empty transcriptions.
• Running headers are not mandatory to transcribe as they are not shown to readers in final texts, but some proofreaders still put them in.

Source and licensing

All text comes from Ukrainian Wikisource, whose text is available under CC BY-SA 4.0. The dataset is released under CC BY-SA 4.0; derivatives must keep the same license and give attribution to Wikisource contributors. The underlying works are, as a rule, in the public domain or otherwise freely licensed according to Wikisource policy, but please check the license of an individual work on Wikisource if this matters for your use.

Each page originates from a Сторінка: namespace page on uk.wikisource.org, whose edit history lists the contributors. You can quickly find the appropriate page and file using Title

Contact

If you noticed any issues with dataset or have any questions, contact me: bicolino34@gmail.com