Vintage-LLM/NCSE

Nineteenth-Century Serials Edition

Historical OCR transcriptions of six British newspapers and periodicals: The Monthly Repository, The Northern Star, The Leader, The English Woman’s Journal, The Tomahawk, and The Publishers’ Circular. Page scans and original OCR come from the Nineteenth-Century Serials Edition (ncse, 2008); layout and reading-order data are credited to J. Bourne, UCL (2025).

Parquet is authoritative: 1,390,519 records, a shared 26-column schema, and 32 shards in six publication directories. Dates range from 1806-01-02 to 1890-12-31. The files include layout boxes, whole-page readings, duplicate-copy records, non-prose and empty placeholders.

Coverage and direct size counts

There are 4,264 issue identities and 93,665 page identities. Issue counts include edition copies, supplements, front/end matter. Direct whitespace word counts are not external tokenizer or pipeline-statistics results.

Publication directory	Shards	Records	Words	Words excluding duplicate boxes	Recommended-mask words
data/English_Womans_Journal/	1	13,375	2,911,014	2,911,014	2,898,791
data/Leader/	6	271,207	62,110,026	37,646,943	36,664,267
data/Monthly_Repository/	2	89,059	17,309,388	17,309,388	17,020,302
data/Northern_Star/	19	904,531	201,419,871	81,288,840	79,980,387
data/Publishers_Circular/	3	100,559	13,651,291	13,651,291	7,260,016
data/Tomahawk/	1	11,788	1,993,252	1,993,252	1,941,962
Total	32	1,390,519	299,394,842	154,800,728	145,765,725

• Parquet: 690,697,431 bytes across 32 files.
• Stored text: 1,771,351,444 Unicode characters, 1,777,315,846 UTF-8 bytes.
• Empty records: 103,059, retained for layout/source-coverage tracking.
• Recommended mask: 736,134 nonempty records, 145,765,725 words.

Words mean len(text.split()); characters mean len(text). manifest.json contains exact per-file hashes/sizes, publication counts and overlapping flag counts. These totals do not represent unique printed content.

Files

Parquet names are part-0000.parquet, etc., under each publication directory. English Woman’s Journal and Tomahawk have one shard each; Monthly Repository has two, Publishers’ Circular three, Leader six, Northern Star nineteen.

data/export.jsonl is a faithful UTF-8 export of the final Parquet: all 1,390,519 records and all 26 columns, no text changes or filters, one JSON object per line in sorted-shard order. Size: 2,839,644,909 bytes. SHA-256:

fd20b17f21ce3997d772b71060d92771f5d2c2da8a71b8490f7d22aadca40b8d

Source scans/PDFs, cached OCR responses, processing tools and tests are not bundled.

Schema — 26 columns

Fields	Types	Meaning
doc_id	string	Unique record identifier.
publication, title_code	string	Display publication name and local configuration/directory code.
issue_id, edition_folder	string	Issue identity and source archive folder label, including copies/supplements.
date, year	string, int32	ISO date and corresponding year.
page_number, reading_order	int64	Source PDF page index (1-based) and within-page layout order. Not necessarily the printed folio.
box_page_id, box_class	string	Layout identity/class; whole-page readings use PAGE / page.
text, n_words, n_chars	string, int64, int64	OCR transcription and exact whitespace word / Unicode character counts.
text_source	string	run, redo, fix or phase2: box OCR, replacement, corrected source-folder reading or whole-page reading.
flags	list of string	Processing, witness, uncertainty and training annotations; flags overlap.
olive_support, olive_recall	nullable float64	Lexical support/recall against original OCR when usable. Scope and quality vary; not transcription accuracy.
v2_agreement	nullable float64	Inherited lexical agreement with the credited upstream reading; null when not recalculated for selected text.
duplicate_of, dup_overlap	nullable string, nullable float64	Main-copy issue identity and inherited issue-copy overlap.
box_dup_overlap, is_duplicate_box	nullable float64, bool	Cross-copy box overlap and derived threshold indicator.
box_xyxy	nullable list of float64	Four pixel coordinates in the credited layout raster; null for whole-page rows.
kcl_issue_dir, kcl_pdf	string	Original source directory/PDF provenance labels, not bundled local paths.

Shards preserve their field types/order and source pandas metadata. Historical processing flags and nullable fields remain available; a flag does not certify that an operation succeeded or that a row is accurate.

Source text and quality review

The principal OCR model was allenai/olmOCR-2-7B-1025-FP8. Additional targeted printed-scan reading and adjudication used gemini-3.8-flash. Original spelling, punctuation, capitals, multilingual quotations and printed inconsistencies are intended to remain. Unreadable spans can be marked [?].

104 printed-scan readings were selected after source review and independent candidate adjudication. They address column interleaving, omitted content, incorrect numerical/table readings and fabricated repetition. 3 additional records have visible HTML table text normalized to tab-separated cells and line-separated rows, retaining their cell content. The final data contain no detected common HTML tags. Other record texts remain byte-identical to their assembled inputs.

Image review was targeted at suspect replacements, large size changes, residual markup, repetition risks and reproducible stratified controls. Available source crops were checked against layout raster dimensions rather than a guessed common DPI. Candidate readings were not automatically accepted merely because they were longer, more fluent, or agreed with another OCR. Several whole-page candidates were rejected for hallucinations or regressions.

Tables can coexist with prose inside a broad box. Newspaper columns are intended to be read top-to-bottom and then left-to-right, not across horizontal strips. Printed repetition is retained when actually present: a reviewed handwriting specimen contains fourteen repetitions of the same sentence. By contrast, an unsupported contents-list loop with 275 repetitions was removed from another record.

Source-backed text reviews do not silently correct an inconsistent printed date/amount on the assumption that the issue date or arithmetic is authoritative. Ambiguous financial/date readings remain explicit review limitations.

Duplicate and non-prose filters

is_duplicate_box targets repeated boxes across different copies/editions, retaining edition-only material. It is derived from duplicate_of and a non-null box_dup_overlap >= 0.5. Excluding all entire copy issues instead (duplicate_of is None) leaves 144,220,226 words, and can discard unique edition material.

Affected box-overlap measurements are recalculated against retained main-copy issue text using casefolded Unicode word 8-grams, NFC/long-s normalization. Issue-copy provenance and inherited issue-level dup_overlap remain unchanged. Matching is approximate, not a semantic-equivalence guarantee. The recommended mask also excludes non_prose, an approximate classification of tables, catalogues, prices, contents/indexes and other primarily list/numeric material. Mixed boxes can contain valuable prose.

Neither mask removes all within-page overlap. Some inherited broad boxes geometrically overlap smaller column boxes. The gross word counts and duplicate-box-filtered counts must not be described as unique printed words.

Agreement scores and remaining limitations

For source-reviewed text, Olive support/recall use the original transparent OCR embedded in the mapped PDF crop, when usable, with Unicode word-count multiset overlap. These recomputed scores have a documented crop/witness scope; they are not interchangeable with every inherited score. Scores are null for unusable witnesses. Upstream agreement is null for changed text rather than carrying an unmeasured stale score. Unreviewed scores remain inherited and can be stale or region-dependent.

Digital witnesses are themselves OCR. Lexical scores ignore word order and do not prove numerical, date or speaker accuracy. This dataset is not certified perfect OCR and has not been manually proofread end-to-end.

Known limitations include:

• Remaining OCR errors, damaged print, partial box boundaries, approximate table formatting, output-cap flags and unresolved source-reading cases.
• Broad/overlapping layout boxes and approximate non-prose/cross-copy classification.
• 82 no_pdf_page placeholders whose page indices lie beyond the available issue PDF. They remain empty; no missing text was invented. Six other selected supplement/end-matter sources could not be reliably mapped in the downloadable PDF bundles.
• Whole-page rows lack layout-box geometry and follow their reading’s internal order.
• Some original witnesses are badly corrupted or misplaced. Genuine Greek/Hebrew and other multilingual print are not errors merely because they disagree with English OCR.
• The eleven English Catalogue of Books volumes remain outside this collection.

Integrity verification

Every written Parquet record and every JSONL object is independently checked: schemas; ids/row order; source geometry/provenance; dates; finite scores; word/character counts; duplicate rules/targets; complete export equality; manifest totals; file hashes/sizes; absence of common HTML, leading YAML and degenerate character runs. All 64 input Parquet files remain unchanged. Processing regression tests and Hugging Face loading were also exercised outside this repository.

Passing structural checks establishes consistency, not exhaustive coverage or universal transcription accuracy. Review uncertainties are documented rather than concealed or repaired by guesswork.

Sources, licence and attribution

This dataset is CC BY 4.0. Please credit:

• Nineteenth-Century Serials Edition (2008), King’s College London, Birkbeck and the British Library: original scans/OCR. Source archive articles: Leader, Monthly Repository, Northern Star, English Woman’s Journal, Tomahawk, Publishers’ Circular.
• J. Bourne, UCL (2025), credited layout/reading-order dataset: figshare article 28381610, CC BY 4.0; associated code: ereading_the_unreadable.

See Creative Commons Attribution 4.0 for the licence terms. Source labels and the review evidence support attribution and further source checking.