Datasets

Danish books (Fraktur)

Twenty complete pages: ten from Søren Kierkegaard’s Af en endnu Levendes Papirer (1838) and ten from Hans Christian Andersen’s Billedbog uden Billeder (1840). The Royal Danish Library digitised both books; Wikimedia Commons supplies the unaltered scans.

Reference text. Validated Danish Wikisource transcriptions, followed by Codex visual comparison of every selected scan. The frozen references preserve historical spelling, case, punctuation, visible word-division hyphens, page numbers and printer’s signatures.

All 20 pages form the complete Danish Fraktur task in the Danish and Scandinavian benchmarks. The additional comparison was performed by an AI assistant, not an independent human transcription team. The five initial models were selected by Swedish Fraktur CER; Swedish inputs are cropped lines, while this task evaluates whole pages.

Danish diplomatic reports

Typewritten Danish diplomatic reports, distributed as the EHRI Danish typewritten line dataset. This collection tests recognition of archival typescript rather than handwriting.

Reference text. Line images paired with text, with source page identifiers and crop coordinates retained in the dataset.

EU Official Journal · Danish

Danish pages from the 2023 Official Journal of the European Union, released in OJ4OCRMT. The benchmark uses the regular-page test subset, rendered at 300 dpi.

Reference text. Raw text extracted from the source PDF with pdftotext. These are digital page renderings with PDF-derived references, not manually transcribed scans.

Historical Danish handwriting

Handwritten city and parish council minutes from 1841–1939, published by Aarhus City Archives. The pages were segmented, transcribed and proofread by archive staff and volunteers.

Reference text. ALTO XML transcriptions. Line images are cropped from annotated boxes; page references join the annotated lines in source order.

Modern Danish handwriting

Contemporary handwriting collected by the Danish National Archives in 2025. Volunteers copied passages from ePAROLE, mainly newspaper and magazine text from the 1990s.

Reference text. Proofread PAGE XML transcriptions. The line benchmark uses all 977 available lines; the page benchmark uses 100 selected pages.

Rectangular line crops can contain parts of neighbouring lines. This affects the line recognition task.

Norwegian handwriting (NorHand)

NorHand contains Norwegian letters and diaries from the 19th and early 20th centuries. The line images capture historical handwriting and spelling; the source release resizes each image to a height of 128 pixels.

Reference text. Transcribed text supplied with each line image.

TrOCR-norhand-v3 was fine-tuned on the NorHand family of datasets; its NorHand result is an in-domain comparison.

Norwegian research documents

Modern Norwegian research documents from Nasjonalt vitenarkiv, the national research repository, including theses and reports. The benchmark selects pages across 19 configured Norwegian PDFs and renders them at 300 dpi.

Reference text. Text extracted directly from each PDF with PDFium. Extraction and reading order can differ from the visible page; references are not independent human transcriptions.

EU Official Journal · Swedish

Swedish pages from the 2023 Official Journal of the European Union, released in OJ4OCRMT. The benchmark uses the regular-page test subset, rendered at 300 dpi.

Reference text. Raw text extracted from the source PDF with pdftotext. These are digital page renderings with PDF-derived references, not manually transcribed scans.

Swedish archival handwriting

Swedish archival text lines from the out-of-domain HTR evaluation release of Riksarkivet (the Swedish National Archives). The benchmark balances its sample across 19 archive and document source files.

Reference text. The transcription supplied with each line image in the source test split.

Swedish newspapers (Fraktur)

Swedish newspaper text in 19th-century blackletter type. Språkbanken produced the transcriptions, and the Swedish National Archives converted the material into line images.

Reference text. Transcribed text supplied with each line image, preserving historical spelling and characters.