NCSE v1.1

The six 19th-century British serials of the Nineteenth-Century Serials Edition (ncse, 2008: King's College London, Birkbeck, British Library), re-read from the original page images with olmOCR-2 (allenai/olmOCR-2-7B-1025-FP8), box by box, using the layout of NCSE v2.0 (Bourne, UCL 2025), with additional whole-page OCR in v1.1. The local Parquet release contains 1,390,028 rows: 1,378,882 layout boxes and 11,146 whole-page records. There are 298,801,318 words, or 154,208,682 words after excluding duplicate boxes (is_duplicate_box, below). Empty rows are retained for layout/coverage tracking.

The dates below describe this release's coverage, not necessarily each serial's entire publication history. issues counts distinct issue_id values, including edition copies, front/end matter and supplements; words_dedup sums n_words where is_duplicate_box is false. Whole-page records are included in these totals.

publication	first	last	issues	rows	words	words_dedup
The Monthly Repository	1806-01-02	1837-12-01	484	88,972	17,269,202	17,269,202
The Northern Star	1837-12-02	1852-11-27	2,215	904,519	201,285,128	81,157,052
The Leader	1850-03-30	1860-11-24	1,005	271,082	61,808,957	37,344,397
The English Woman's Journal	1858-03-01	1864-08-01	91	13,375	2,910,785	2,910,785
The Tomahawk	1867-05-11	1870-08-27	187	11,711	1,985,875	1,985,875
The Publishers' Circular	1880-01-17	1890-12-31	274	100,369	13,541,371	13,541,371

Local files and loading

The full dataset is the 32 Parquet shards under data/, all exposed as a train split; there are no separate validation or test splits. The default configuration covers all six publications; each named configuration in the YAML header selects one publication directory.

directory	Parquet files	rows
data/English_Womans_Journal/	part-0000.parquet	13,375
data/Leader/	part-0000.parquet through part-0005.parquet	271,082
data/Monthly_Repository/	part-0000.parquet through part-0001.parquet	88,972
data/Northern_Star/	part-0000.parquet through part-0018.parquet	904,519
data/Publishers_Circular/	part-0000.parquet through part-0002.parquet	100,369
data/Tomahawk/	part-0000.parquet	11,711

Other release files:

• README.md: dataset description and loading configurations.
• STATS.md: publication, edition, row-class, flag and text-source statistics.
• manifest.json: v1.1 release counts, dates, duplicate-filtered word totals and shard counts.

For example, with the Hugging Face datasets package installed:

from datasets import load_dataset

data = load_dataset(
    "parquet",
    data_files={"train": "data/*/*.parquet"},
    split="train",
    streaming=True,
)
training_text = data.filter(
    lambda row: row["n_words"] > 0 and not row["is_duplicate_box"]
)
print(next(iter(training_text))["text"])

To select one publication, replace the glob with, for example, data/Northern_Star/*.parquet.

New in v1.1: pages NCSE v2 has no boxes for

NCSE v2 (and so v1) has no layout boxes for some pages that the 2008 OCR shows to have text: 127 regular issues of the Monthly Repository (1806-1836) plus its end matter and supplements, 38 issues of the Publishers' Circular, and single pages elsewhere. v1.1 reads these whole pages with the same model (cut at the column gutters and into pieces that fit the model) and adds them as rows with box_page_id = PAGE, box_class = page, text_source = phase2 (no v2 layout, so no reading order inside the page beyond the model's own). A page whose text is already in v1 (>= half of its word 8-grams in a copy of the same date) is emptied and flagged page_in_v1. The 11 volumes of The English Catalogue of Books (bibliographic listings) are not included.

publication	pages	pages_with_text	words	already_in_v1	median_olive_support
The English Woman's Journal	25	24	10,510	1	0.930
The Monthly Repository	9,366	9,248	5,458,809	97	0.953
The Northern Star	3	0	0	3
The Publishers' Circular	1,695	1,411	1,213,611	283	0.904
The Tomahawk	57	46	23,406	3	0.123

The Tomahawk's phase-2 pages are mostly illustrated advertisement pages. On the pages checked, the 2008 OCR of them is unreadable (strings like ^^ g ^ sg I" iscaaar), so their low olive_support reflects the witness, not the text.

Why a re-OCR

NCSE v2.0 transcribed every layout box with Pixtral-12B. Checked against the independent 2008 OCR of the same pages (Olive Software, word level with coordinates), v2 has repetition loops (about a quarter of its words are in rows whose 6-grams repeat), invented text in capped rows, and descriptions written in place of transcriptions ("The image is a page from a 19th-century English newspaper…" on whole advertisement pages). v1 keeps v2's layout boxes and reading order but reads every box again from the KCL page images. v2's text is never used as text.

Against the 2008 Olive OCR (same box matching for both; boxes with a usable witness): support = share of the transcribed words that Olive also has (low = invented or looping text), recall = share of Olive's words that the transcription has (low = text missing). Scores of the full run, before the post-pass that redid 9,488 truncated boxes (which raised the matched words of those boxes from 2.52 M to 3.24 M). These are reported production-run comparisons, not scores independently recomputed from this local package.

publication	support v2	support v1	recall v2	recall v1
The Tomahawk	0.812	0.948	0.743	0.788
The English Woman's Journal	0.920	0.942	0.919	0.923
The Monthly Repository	0.918	0.954	0.935	0.946
The Publishers' Circular	0.820	0.920	0.814	0.800
The Leader	0.816	0.932	0.783	0.900
The Northern Star	0.750	0.881	0.723	0.871

Fields

All shards share the same 26-field schema.

fields	meaning
doc_id	Record identifier.
publication, title_code	Display title and local publication/configuration code.
issue_id, edition_folder	Issue identifier (as in NCSE v2) and source KCL edition-folder label.
date, year	Issue date as an ISO YYYY-MM-DD string and integer year.
page_number, reading_order	Integer page number and record order within the page.
box_page_id, box_class	Layout-box ID and class (text, title, table, figure); whole-page additions use PAGE and page.
text, n_words, n_chars	OCR transcription and integer word/character counts; empty records have empty text and zero counts.
text_source	run = full box OCR run; redo = re-OCR'd with more splitting; fix = issue re-OCR'd from the right folder; phase2 = whole-page OCR added in v1.1.
flags	List of processing/quality flags; may be empty.
olive_support, olive_recall	Nullable scores against the 2008 OCR when it is a usable witness (defined in “Why a re-OCR”).
v2_agreement	Nullable share of words also present in the NCSE v2 transcription of the box.
duplicate_of, dup_overlap, box_dup_overlap, is_duplicate_box	Nullable main-copy issue ID and copy/box overlap scores, plus a Boolean duplicate-box indicator (see below).
box_xyxy	Nullable list of four pixel coordinates [x0, y0, x1, y1] in the v2 page image; null for phase-2 whole-page rows.
kcl_issue_dir, kcl_pdf	Source-archive issue directory and PDF filename used for OCR.

edition_folder, kcl_issue_dir and kcl_pdf are source provenance, not paths to files bundled here. In particular, source labels such as English_Woman's_Journal differ from the local directory data/English_Womans_Journal/; the edition folders listed below are also archive labels, not local subdirectories. The original page PDFs/images and the 2008 Olive OCR are not included in this folder; use the credited source archives.

How the text was made

• Each v2 box is cropped from the KCL page PDF. Boxes wider than one column are cut at the column gutters (gutters found per band of the image, or taken from the page's own column boxes); tall or dense boxes are cut into pieces that fit the model's context. Pieces are joined in order; a word hyphenated across a piece break is rejoined.
• Prompt: transcribe exactly as printed (spelling, punctuation, capitals, hyphenation), no modernising, no correction; an illegible word is written [?]. These markers are kept: they mark where the page could not be read.
• Cleaning, each logged in flags: repeating tails of crops that hit the 3,500-token cap cut (loop_cut); HTML tables turned into one line per row, cells tab-separated (table_from_html); stray markup removed (markup_stripped); dot/underscore leaders shortened to five (leaders_shortened); boxes with only [?] emptied (illegible_only); drop capitals joined to their word (dropcap_*); boxes that came out mostly in CJK/Thai/ Devanagari script, never printed in these papers, emptied (hallucinated_script); words where a Latin letter is glued to Cyrillic/Arabic/... letters flagged, not changed (mixed_script_word). Greek and Hebrew quotations (mostly in the Monthly Repository) and real Arabic are kept as printed.
• Slivers (v2 boxes under 25 px high and 8x wider than high: seams between two boxes, letters cut in half) are emptied (sliver_dropped): their text was mostly invented (only 47-61 % supported by the 2008 OCR), and the neighbouring boxes read most of the same line.
• Witness flags: low_olive_support (usable 2008 witness, support < 0.5), low_v2_agreement (< 0.5 of the words also in v2; often v2 described the box), unconfirmed (both: treat with care), still_capped, ocr_error, no_text.

Repeated copies of the same issue

The archive keeps every bound copy the British Library had: the Northern Star in up to nine editions of the same week, the Leader in town and country editions. Later editions repeat most of the paper and add late news. duplicate_of names the main copy of the same date (the KCL main-run folder, else the longest copy) when at least half of a copy's sampled word 8-grams are in it; box_dup_overlap is the share of a box's 8-grams found in the main copy; is_duplicate_box = in a copy and overlap >= 0.5 (the histogram dips between 0.1 and 0.4). For training use not is_duplicate_box (154,208,682 words): it keeps the main copies and the edition-only boxes. This is different from dropping entire copies with duplicate_of is not None, which leaves only 143,660,303 words.

Known limits

• v1 lacked pages without NCSE v2 layout boxes. v1.1 adds 11,146 whole-page records, as detailed above; these have no v2 box geometry or box-level reading order. Of these, 417 rows are empty, including 387 flagged page_in_v1 because their text was already covered. The three Northern Star additions are all empty for that reason. The 11 English Catalogue of Books volumes remain excluded.
• The 2008 witness is itself OCR and is sometimes misplaced on the page; where olive_support is low but v2_agreement is high, the two independent model readings agree and the witness is the outlier.
• Box-row reading order is v2's; phase-2 whole-page text follows the model's own reading order. Table layout is approximate.

Checks

The release documentation reports an independent verifier checking that every v2 box (and in v1.1 every phase-2 page) is present exactly once; counts, dates, sorting and duplicate fields are consistent; there is no markup or copied v2 description text; and 400 random untouched boxes per publication were re-derived from raw OCR output. Its reported result was PASS. The raw OCR inputs are not bundled, so that full production-run verification cannot be independently reproduced from this local package.

Local checks of all 32 Parquet shards confirm the shared schema and all counts, dates and duplicate-filtered totals in manifest.json, the publication, edition, phase-2, row-class and flag tables in this README, and all numerical tables in STATS.md. Every data-file glob in the YAML header resolves to the expected shards.

Sources, licence, credit

• Page images and the 2008 OCR: ncse: Nineteenth-Century Serials Edition (2008), King's College London, Birkbeck, British Library; archive on KCL figshare (Leader 25103786, Monthly Repository 25108519, Northern Star 25106895, English Woman's Journal 25109275, Tomahawk 25109047, Publishers' Circular 25101957), CC BY 4.0.
• Layout boxes, reading order, classes: NCSE v2.0, J. Bourne (UCL, 2025), figshare 28381610, CC BY 4.0; code github.com/JonnoB/ereading_the_unreadable.
• OCR model: allenai/olmOCR-2-7B-1025-FP8. This dataset: CC BY 4.0. Please credit ncse (2008) and NCSE v2.0 as above.