lector-bench v0 (preliminary)

Page-level diplomatic transcription of Latin scholastic print, 1602–1907. Each item is one page scan and its transcription as a Markdown page file: the text exactly as printed, with its footnotes, marginal notes and notes carried over from the previous page in place. The task tests what page readers (OCR engines, vision-language models) actually fail at on this material: dense two-column early-modern pages, abbreviations and ligatures, marginalia, and footnote apparatus.

v0 is preliminary. Its reference transcriptions were made by model pipelines and checked by models, not yet by a human (see How the references were made). Use it to compare systems, not to certify one. A human-verified v1 will follow.

Contents

work	year	pages	what it tests
Diego Álvarez, De auxiliis (Rome)	1610	30	dense two-column print, abbreviations, marginal citations and paragraph summaries
Domingo Báñez, Scholastica commentaria	1602	6	heavy contractions, two columns, small type
Vincent Contenson, Theologia mentis et cordis	1681	5	early-modern print
Pedro Hurtado de Mendoza, Universa philosophia (Lyon)	1624	10	old print with marginal summaries
Salmanticenses, Cursus theologicus I	1870	5	modern reprint, two columns
Vincent Contenson (1875 edition)	1875	5	modern reprint
J. B. Franzelin, De Ecclesia Christi	1887	5	footnotes, notes continued across pages
Louis Billot, De inspiratione / De immutabilitate traditionis	1903, 1907	20	footnotes, continued notes

manifest.csv lists every page with its work, year, edition, scan source, print era, track, provenance and features (footnotes, marginal notes, paragraph summaries, continued note, length).

Tracks. In-domain pages come from works whose other pages trained the lector readers; out-of-domain pages come from works no lector model saw. A hidden split of modern editions (which cannot be redistributed) is kept privately and reported separately.

The page-file format

• One paragraph per block (blank line between); *italic*, **bold**; headings as whole-bold lines.
• Footnotes as ^[text] right after their anchor.
• Printed marginal notes as ^[⟦M⟧ text] at the passage they stand beside; a marginal summary of a whole paragraph as ^[⟦M⟧⟦R⟧ text].
• A footnote continued from the previous page as a last block starting ⟦CONTINUED-NOTE⟧.
• Diplomatic: the print's spelling, æ/œ, &, accents, u/v, i/j and abbreviation marks are kept; long s is written s. Printer's errors are kept as printed.
• Running heads, page numbers, signatures and catchwords are omitted.

Scoring

score.py (needs rapidfuzz) compares a directory of predictions laid out like data/:

python score.py data runs/my-system --fold

• md_cer: edit distance over the whole page file (text, markup, notes) / reference length. Where a marginal note sits inside its paragraph is a convention, not something the page shows, so it isn't scored.
• cer_body, cer_notes, wer: text only.
• found: notes recovered, by kind (footnotes, marginal notes, continued notes).
• omit30 / insert30: pages where the reader skipped a run of 30+ words (a column jump) or produced 30+ words the page doesn't have (a loop, or text recited from memory). These are how vision-language models fail where classical OCR doesn't, and a mean CER hides them.
• --fold folds orthography (æ→ae, j→i, accents) for systems that don't transcribe diplomatically.

Baselines (v0, single page at 1280 px longest side, greedy decoding)

Public split, 86 pages, score.py --fold. md_cer is the page-weighted mean; omit30 and insert30 count pages where a run of 30+ words was skipped or invented. Outputs pass through one loop guard (a chunk of 12+ characters repeated 3+ times in a row is cut to one copy), as the deployed reader does.

Model	md_cer	omit30	insert30	marginal notes found
lector ft2 (Qwen3-VL-2B, full fine-tune; internal)	0.059	8	1	4 / 103
lector ft4 (Qwen3.5-2B; internal)	0.089	14	3	45 / 103
lector-2b-v1 (Qwen3.5-2B, faithful targets)	0.066	6	3	48 / 103
lector-4b-v1 (Qwen3.5-4B, same data)	0.039	3	1	71 / 103

The errors concentrate where a mean hides them: Báñez 1602, dense and heavily abbreviated, which no lector training page resembles (4b-v1 0.26, 2b-v1 0.45), against 0.002–0.04 on the 19th–20th-century works. 2b-v1 also loops on a page with Greek (Salmanticenses, 0.55 on that page); 4b-v1 drops the second column of one Hurtado page. No general-purpose or commercial system is scored yet.

How the references were made

The transcriptions come from the Schola Thomistica digitization pipeline (scriptorium), under its diplomatic page-file contract. They were drafted by model readers, then corrected and adjudicated by other models on crops of the scan:

• Álvarez: LLM vision readers, an independent LLM critic, adjudication on the scan.
• Billot and Hurtado: drafted by lector ft4, corrected by DeepSeek without human edit. These pages flatter lector models, whose errors a corrector may have left in place.
• Báñez 1602 and Contenson 1681: read by Claude Opus on native-resolution quarter crops with a self-check pass; words where two independent systems disagreed with the reference were re-read on the image.
• The 1870–1887 pages: model readings checked against the scan.

For v0 each new page was also checked against Apple Vision OCR for missing lines. Words where two independent readers (lector ft2 and ft4, or ft2 and Apple Vision) agreed against the reference were re-read on the image by Claude: 182 such words on the 55 new pages, of which about 46 were reference errors (misread words, expanded abbreviations, normalized spellings) and were corrected; the rest were reader errors. One Hurtado page (p. 64), where most flagged words were reference errors, was re-read whole from the scan. Corrections are listed with the code (bench/corrections.tsv in the lector repository). Errors that neither reader caught remain possible, more so on the Hurtado pages, whose draft came from a lector model. Emendations the editors made (the print reads X; read Y) are reversed, so the reference keeps the print's reading.

License and sources

The transcriptions are released under CC BY 4.0. The page images reproduce printed books in the public domain (published 1602–1907; the latest author, Louis Billot, died in 1931). A faithful scan of a public-domain page adds no copyright of its own (Bridgeman v. Corel, 1999; EU Directive 2019/790, art. 14), so the images are shared as public domain. Each image's source is in the manifest (Google Books, the Bayerische Staatsbibliothek, the Internet Archive); please credit it. If you hold a contrary claim, open a discussion on this dataset's page and the image will be reviewed.

Citation

Preliminary; a paper is in preparation. Until then, cite this dataset by its Hugging Face page.