Ericu950/litteraturbanken-ocr-corpus

Litteraturbanken OCR Corpus

Machine-read text of Litteraturbanken's facsimile collection, produced by a fine-tuned vision-language OCR model (see Ericu950/swedish-print-ocr-3b). 13,647 works, 2,260,417 pages, 2.95 billion characters, one row per printed page, keyed by the catalogue's own page name. 179,768 of those pages (247.6 million characters) predate 1800, including the 1541 Gustav Vasa Bible in full and all five volumes of Rudbeck's Atlantica (1679–1702).

This is the only copy of this release build. One JSONL file per book lives under books/, plus metadata.jsonl / metadata.tsv with one row per book. Because the Hub caps directories at 10,000 files, books/ is sharded by the two characters that follow lb in the work id: lb16abc….jsonl lives at books/16/lb16abc….jsonl. The release is a filter over a larger internal corpus (far_spridas == "ja"), not a separate copy of it — see Licence handling below.

Schema — one row per page

{"sidnamn": "47", "sida": 51, "text": "...", "far_spridas": "ja"}

Field	Meaning
sidnamn	the page's name in Litteraturbanken's catalogue — 47, XII, -3
sida	the catalogue's imagenumber, or the PDF page for PDF-read volumes
text	the transcription
far_spridas	ja or nej, see Licence handling
ofullstandig	loop or avklippt (truncated) if present; absent means the page is complete

Rows are in catalogue order. sidnamn is the key, not the row number — volumes with pages inserted after the fact have correct order but not contiguous row numbers against any earlier build.

Licence handling

Litteraturbanken publishes under several licences, and the distinction is legal, not technical.

• ja for the whole work — pd, cc-0, cc-by.
• Per page — lb-svs, lb-assv, lb-sls (243 works). These licences state that the editor's introduction and commentary may not be redistributed while the literary text is CC BY. The boundary is drawn page-by-page by a language model (Qwen3.6-27B) reading the first 50 words of each page without being told the book's title — metadata (parts[].texttype) was not reliable enough: one volume of textual commentary was misclassified as free text because its fourteen kapitel-typed sections are named after the works they comment on. Of 78,200 pages in these works, 43,892 (56%) are the author's own.
• Never released — 172 works whose licence outright forbids redistribution are not read at all and are not in this corpus. An index of what was excluded ships alongside (see the source project's EJ_SPRIDNING/INDEX.jsonl, not included in this upload).

What's verified

Page names against Litteraturbanken's own proofread e-text. For 164 works LB also publishes a proofread e-text keyed by the same page name. Comparing this corpus's page n against the e-text of pages n−2 … n+2 by character-trigram Jaccard gives 0.949 at the correct page against 0.258 at its neighbour — 144 of 145 testable volumes have their peak in the right place. The one exception, lb1963958, is Litteraturbanken's own: the facsimile catalogue and the e-text number that volume two pages apart (confirmed against the image); this corpus follows the facsimile catalogue.

Page counts against the catalogue, per volume. sidor and sidor_i_katalogen sit side by side in the metadata. 301 pages are missing corpus-wide, and Litteraturbanken has no image for them either.

Before using this text

• Nothing is proofread by a human.
• Page numbers, running heads and marginal notes are absent by instruction — readable, but not citable without sidnamn.
• Normalised, not diplomatic. Long ſ is written as s.
• tryckt: false marks manuscripts, mostly medieval (e.g. Karlskrönikan c. 1450, Själens tröst c. 1430). The model is trained on print and reads these only partially — recognisable but heavily corrupted Old Swedish. Kept because they're free and someone may want them, but should not be mixed with print-OCR quality claims.
• A spread counts as one page. Some facsimiles photograph two book pages at once — the 1541 Bible averages 6,127 characters per catalogue page against 1,127 for a normal 19th-century volume.
• Parallel-text editions sometimes lose a column. Rudbeck's Atlantica has Swedish fraktur on the left and Latin on the right; 72–84% of pages contain Latin, but not all — this could not be prompted away (three instruction variants, two explicitly requesting both columns, returned the same 1,678 characters on the same test page).

Attribution

This material comes from Litteraturbanken.se and the libraries that digitised it. When using it, credit Litteraturbanken and the library named for the relevant work.

Related

• Model that produced this corpus: Ericu950/swedish-print-ocr-3b
• Held-out benchmark model: Ericu950/swedish-print-ocr-3b-benchmark