Vintage-LLM/HMD-LWM

HMD + LwM British Newspapers

Every article of the two openly licensed British Library newspaper collections — Heritage Made Digital (HMD, 14 title IDs) and Living with Machines (LwM, 107 titles) — in one dataset: 9,326,170 articles, 5,705,878,113 words, 121 titles, 1801-1920 (7,185,653 articles / 4,858,146,473 words up to 1900). Text is the British Library's own OCR (Nuance/Kofax OmniPage 20 in CCS docWorks, run by Findmypast), extracted with the Living with Machines alto2txt tool, verified against every other public copy, then layout-repaired without changing a single letter or digit. Near-duplicates are marked, not removed. Every row carries a trace back to its original in each online source.

What is this

• v0 = the original BL plaintext (alto2txt), byte-identical in all public copies (see Verification). It is not re-shipped here; every row has text_raw_sha256 and exact locations to re-fetch it.
• v1 = layout repair only: OCR lines joined into paragraphs, end-of-line hyphenation rejoined (using the OCR engine's own hyphenation marks from the ALTO XML where available, a dictionary compound rule everywhere), whitespace normalised. Invariant checked on every row: the sequence of letters and digits is identical to the original. No spelling correction, no long-s repair, no case changes, no deletions, no model-generated text.
• The OCR is often poor (commercial OCR on 19th-century print; many titles scanned from microfilm). Use pwer, quality_band and bl_ocr_quality_mean to select.

Collections

collection	titles	first_year	last_year	articles	words	words_le_1900
HMD	14	1801	1871	2,901,554	2,277,956,708	2,277,956,708
LwM	107	1802	1920	6,424,616	3,427,921,405	2,580,189,765

Quality

pwer is a predicted word error rate (a linear model on subword fertility, dictionary hit rate and short-junk share, fitted on 193 hand-corrected 19th-c. American newspaper articles; bands A < 0.20 <= B < 0.50 <= C). It is a routing signal, not a measure of faithfulness. bl_ocr_quality_mean is the OCR engine's mean word confidence.

band	articles	words	pct_words
A	3,600,234	2,772,513,161	48.59
B	3,108,742	2,408,776,190	42.22
C	2,478,308	524,588,762	9.19
none (no words)	138,886	0	0

By decade (word-weighted):

decade	articles	words	pwer_w	A_pct	B_pct	C_pct
1800	154,083	141,017,885	0.365	9.4	74.4	16.2
1810	214,413	189,609,233	0.348	17.1	65.9	16.9
1820	260,586	237,387,562	0.237	48.5	47.6	3.9
1830	191,654	181,941,000	0.199	69.3	27.5	3.2
1840	602,847	595,057,014	0.205	65.7	31.2	3.1
1850	989,058	661,273,285	0.246	53.2	39.3	7.5
1860	989,080	645,246,632	0.225	59.5	35	5.5
1870	801,849	535,269,693	0.262	49.1	40.8	10
1880	1,346,118	862,396,945	0.257	48.4	43.3	8.3
1890	1,493,709	742,865,528	0.281	44	44.4	11.6
1900	1,349,945	575,419,463	0.305	40.6	43.7	15.6
1910	868,159	315,619,079	0.312	34.6	50.7	14.7
1920	64,669	22,774,794	0.297	37.6	48.8	13.5

Article sizes

Many "articles" are whole pages or advertisement pages where the original article segmentation failed.

size	articles	pct_articles	pct_words
a <20	1,700,096	18.2	0.2
b 20-99	1,537,750	16.5	1.5
c 100-499	3,223,669	34.6	14.4
d 500-999	1,249,448	13.4	15.6
e 1k-5k	1,486,086	15.9	50.6
f >=5k	129,121	1.4	17.7

Near-duplicates

MinHash over word 5-grams (64 permutations, LSH 16x4), estimated Jaccard >= 0.6, articles of >= 20 words, and different papers only within 60 days of each other. One representative per cluster (dup_is_representative), chosen by best dictionary hit rate, then lowest fertility, then earliest date. Taking only representatives keeps 5,543,814,781 words.

kind	articles	redundant_articles	redundant_words
cluster across papers	252,800	153,341	70,658,489
cluster within one paper	257,429	168,712	91,404,843
unique	8,815,941	0	0

Verification (all rows, not samples)

• BL repository plaintext: 1,264 zips, every one checked against the repository's CRC64NVME checksum.
• Zenodo CSV releases (10.5281/zenodo.15056046, 10.5281/zenodo.17425252): all 9,326,170 articles byte-identical.
• LwMDB open item index: same 9,326,170 article ids and word counts.
• HF biglam/hmd_newspapers: all 3,065,408 rows are articles of this dataset (63 rows of one Northern Guardian issue carry a date one day earlier there: hf_biglam_date_minus_1).
• HF Livingwithmachines/hmd-erwt-training (269,136 rows; the Hub viewer shows 134,222 because it stops at 5 GB): all its 50,242 issues present, page-level cut of the same OCR: 96.8 % of our whitespace tokens appear in it per issue; on the worst-agreeing issue inspected (The Statesman, 1814-10-01) page word counts, confidences and the letters+digits sequence are identical — that dataset puts a space between every OCR token.
• Independent re-extraction from the BL METS/ALTO XML (894 zips, 138.7 GB, reproduces alto2txt): byte-identical for 6,408,772 articles; see alto_status.
• A second, independently written verifier re-read every BL zip: PASS, 9,326,170 rows = 9,326,170 zip members, no failures, Zenodo sha256 equal on 9,326,170.
• Re-checked on 2026-09-28: every row's original text equals its BL member (sha256, byte length, letters+digits), every member of the 1,264 zips is used once, the title ids equal the Zenodo releases (14 HMD, 107 LwM), 15,531 articles re-read from Zenodo are identical, and every table in STATS.md and stats/ was recomputed from the parquet files. How each row was made and how to re-fetch its original: PROVENANCE.md.

alto_status	articles	pct_words
match	6,408,772	77.19
no_xml_for_year	1,555,071	11.02
no_xml_for_title	1,243,213	10.31
no_alto_row	113,536	1.41
differs	5,578	0.07

Known defects of the source

• Article segmentation (OLR) often failed: page-sized items, adverts mostly tagged ARTICLE.
• The BL METS/ALTO XML release is incomplete: no XML for 27 LwM titles; essentially none after 1900; St Helens Examiner's XML zips are partial; Colored News 1855-08-04 page 3 ALTO is a stray text note; some zips contain partial .azDownload-* files. The plaintext (this dataset's source) is complete.
• Some LwM titles are Welsh-region papers; a language tag is not provided — detect it yourself.

Fields

See PROVENANCE.md for every column. Key ones: doc_id (<NLP>_<YYYYMMDD>_artNNNN), collection, nlp (BL title id), newspaper_title, location, date, item_type, text, n_words, pwer, quality_band, bl_ocr_quality_mean, dup_cluster_id, dup_is_representative, text_raw_sha256, bl_item_doi, bl_zip, bl_member_txt, zenodo_doi, zenodo_file.

Official sources

• British Library Research Repository, https://bl.iro.bl.uk: the alto2txt plaintext items (one DOI per title group, bl_item_doi, e.g. https://doi.org/10.23636/ej4h-6236) and the METS/ALTO XML items. Files: https://bl.iro.bl.uk/server/api/core/bitstreams/<bl_bitstream_uuid>/content.
• Zenodo CSV releases of the same text: HMD https://doi.org/10.5281/zenodo.15056046, LwM https://doi.org/10.5281/zenodo.17425252. LwMDB item index: https://doi.org/10.5281/zenodo.15102269.
• Hugging Face copies: https://huggingface.co/datasets/biglam/hmd_newspapers, https://huggingface.co/datasets/Livingwithmachines/hmd-erwt-training.
• alto2txt: https://github.com/Living-with-Machines/alto2txt (Living with Machines code: https://github.com/Living-with-Machines).

Licence and credits

The newspapers are in the public domain; the BL repository publishes the text under CC0 1.0, and this derived dataset is released under CC0 1.0. (The Zenodo CSV copies are published under CC BY 4.0; the text here comes from the CC0 BL items.) Digitisation: British Library (Heritage Made Digital; Living with Machines, AHRC grants AH/S01179X/1 and AH/Y00745X/1) with Findmypast. Text extraction: Living with Machines alto2txt. Please cite the BL repository items (bl_item_doi) and the Living with Machines project.