Vintage-LLM/HMD-LWM

HMD + LwM British Newspapers

British Library newspaper OCR from Heritage Made Digital (HMD) and Living with Machines (LwM), restricted to 1801-01-01 through 1913-12-31.

• 8,716,132 archive articles, 120 digitisation title IDs, 113 annual Parquet files.
• 5,498,203,322 full-text words and 4,976,443,276 heuristically retained prose words.
• 5,376,959 training articles / 4,487,951,298 training words.
• Every final training string is byte-distinct. Archive records are not deleted when training copies are deduplicated.
• 61 columns; source text and source-retrieval metadata are preserved. No generated replacement text or image-based re-OCR.

Contents and loading

data/1801.parquet through data/1913.parquet contain the archive, not just the training selection. manifest.json records totals, schema, per-year statistics, file sizes and full-file SHA-256 checksums. PROVENANCE.md defines the fields and selection policies; VERIFICATION.md gives the checks, evidence scope and limitations. All local paths are relative to this directory.

Filter the Boolean train column explicitly. A loader split named archive contains every archive row; a split name does not apply the training filter. Use coalesce(text_prose, text) for training, not unconditionally text.

import duckdb

con = duckdb.connect()
con.execute("""
    CREATE VIEW training AS
    SELECT doc_id, coalesce(text_prose, text) AS text
    FROM read_parquet('data/*.parquet')
    WHERE train = true
""")
print(con.execute('SELECT count(*) FROM training').fetchone()[0])
# 5376959

Streaming with PyArrow:

import pyarrow.dataset as ds
import pyarrow.compute as pc

archive = ds.dataset('data', format='parquet')
scanner = archive.scanner(
    columns=['doc_id', 'text', 'text_prose'],
    filter=ds.field('train') == True,
)
for batch in scanner.to_batches():
    training_text = pc.coalesce(batch.column('text_prose'), batch.column('text'))
    # Consume training_text without loading the whole corpus into memory.

Selection and text policy

train is true exactly when the row is band A/B, the retained near-duplicate representative, guessed English, has at least 20 retained-prose words, and has an empty flags list. Flags include inherited quality-gate exclusions, additional lexical/repetition review, a conservative all-short numeric-fragment screen, and exact final-training-text duplication.

The selection funnel is: 8,716,132 → 6,277,988 → 6,002,144 → 5,481,690 → 5,445,913 → 5,376,959 (archive → A/B → representatives → English → 20+ prose words → no flags).

text is the layout-reflowed BL OCR, not a corrected transcription. text_prose is a reversible paragraph-cut view: offsets in non_prose_spans identify what was removed. A null prose value means no cuts; an empty string means all text was cut. Narrative adverts, verse and dialogue may remain. The prose label, language guess and quality scores are heuristics, not adjudicated ground truth. Reviewed narrative passages are retained verbatim; unclear OCR is not rewritten into plausible English. See PROVENANCE.md for exact interpretation of the fields.

Collections

Collection	Titles	Years	Articles	Full-text words	Retained-prose words	Training articles	Training words
HMD	14	1801–1871	2,901,554	2,277,956,708	2,059,795,068	1,655,982	1,869,290,013
LwM	106	1802–1913	5,814,578	3,220,246,614	2,916,648,208	3,720,977	2,618,661,285

Quality

Band A: predicted word error rate <0.20; B: 0.20–<0.50; C: ≥0.50. Rows with zero canonical words have no band. These are inherited model signals, not measured transcription accuracy against page images. Band C is excluded from training but remains in the archive.

Band	Articles	Full-text words	Retained-prose words	Training words
A	3,387,521	2,704,107,694	2,641,571,814	2,499,544,494
B	2,890,467	2,303,368,626	2,022,672,380	1,988,406,804
C	2,308,192	490,727,002	312,199,082	0
none	129,952	0	0	0

Language guesses

Predominantly English, with cheap function-word language guesses. Many non-English labels, especially on poor OCR, can be false positives. Nothing labelled other than English is admitted to training. Labels below 15 canonical prose words are unknown. Two independently read English-verse items are explicitly English; other uncertain labels are not bulk-converted.

Guess	Articles	Retained-prose words
cy	55,989	4,559,567
de	251	9,645
en	6,505,532	4,964,578,557
fr	249	8,577
gd	847	29,471
la	3,035	265,488
unk	2,150,229	6,991,971

Counting and limitations

All word totals above count lowercase ASCII [a-z0-9]+ tokens. Thus a number such as 4,650,352 counts as three words, and don't or Hanover-square counts as two. Counts are not whitespace-word counts or language-model tokens.

Source article segmentation can merge stories, tables and adverts into page-like items. Reading order, OCR spelling, missing punctuation and dehyphenation can still be wrong. The prose and numeric screens are imperfect; training eligibility is not a guarantee of clean prose. No page images were consulted for this build. Public copies of the same OCR can agree while sharing transcription errors. Historical source-quality/gate programs are not bundled, and the original gate was not rerun. No external-index feature statistics are presented as a census.

Licence and attribution

Source plaintext is released by the British Library Research Repository under CC0 1.0. This BL-plaintext-derived package is released under CC0 1.0. Separate Zenodo copies have CC BY 4.0 terms; they are corroborating witnesses, not the licensing basis asserted for this package.

Credit the British Library's Heritage Made Digital and Living with Machines projects (AHRC AH/S01179X/1 and AH/Y00745X/1), Findmypast digitisation/OCR, and the Living with Machines alto2txt extraction project. Cite the source repository items identified by each row's bl_item_doi.