Hansard-v1: UK Parliamentary Debates, 1803-1913

UK parliamentary debate sections and related sitting, appendix and front-matter records from the printed Hansard, covering 1803-1913, with the gaps described below. The release contains 231,313 rows and 409,062,915 words, of which 211,876 rows and 403,402,762 words are marked as prose (is_prose = true: the recommended training subset). The actual earliest and latest record dates are 1803-03-29 and 1913-12-12; 1913-12-31 is the collection cutoff, not the date of the final record.

Files and loading

• README.md: dataset description, schema and loading configuration.
• manifest.json: release totals, date range, per-shard row counts, series statistics and assembly logs. assemble_stats includes intermediate processing counts, which need not equal the final retained-row counts.
• verify.json: recorded independent verification results, including coverage and sampled source checks.
• data/: 13 Parquet shards in 12 decade directories, detailed in the size table below. Each directory contains part-0000.parquet; data/1900s/ also contains part-0001.parquet.

The YAML configuration loads all shards as a single train split. There are no separate validation or test splits. With the Hugging Face datasets package installed, run from the dataset root:

from datasets import load_dataset

data = load_dataset(
    "parquet",
    data_files={"train": "data/*/*.parquet"},
    split="train",
    streaming=True,
)
prose = data.filter(lambda row: row["is_prose"])
print(next(iter(prose))["text"])

To load only one decade, use a pattern such as data/1880s/*.parquet.

Source

• Official Hansard archive XML (hansard-archive.parliament.uk): the Digital Hansard project's per-volume XML, one zip per volume. This is the primary text; speeches are tagged with the speaker, sections with titles, sitting dates and column numbers.
• Historic Hansard website (api.parliament.uk/historic-hansard, formerly Millbank): the same XML rendered as web pages. Used only for what the archive zips lack: whole volumes missing from the archive, volumes whose zip holds only front matter, and single columns missing from a zip. 75,885 sections, 64.0 M words (source = historic_hansard_web).
• The text is the original OCR of the printed volumes as published by Parliament, lightly corrected by the project that made it; nothing was re-OCR'd or rewritten here. The production notes report predicted-WER sample scores at the scorer's floor (~0.10) in every series. These scores were not recomputed from this package and are not measured error rates. Occasional OCR slips remain (e.g. "OP" for "OF").

The original XML ZIPs, saved source pages and column-index inputs are not bundled. Their names and URLs in the records identify external sources, not files or folders that should exist under data/.

What was done

• Debate rows represent top-level sections; other rows represent sitting-level text, appendices or front matter (part, below). Nested sub-section titles stay inside the text as their own paragraph. Speeches read as printed: "Mr. Cobden rose to bring forward his Motion...". Paragraphs are separated by a blank line.
• Dropped (not prose): division lists (who voted Aye/No), tables, indexes, tables of contents, title pages. Counts per row: tables, divisions, dropped_words.
• Words broken across a column or page ("Gal- / lant") are rejoined when the joined form is at least as common in the corpus as the hyphenated one (100,221 joined, 2,225 kept hyphenated, e.g. "to-morrow", "time-honoured").
• Dates: the XML has typos ("1101-08-10", "1991-05-17"); a date more than 400 days from its volume's median takes the previous good date (flag date_repaired, 423 rows); appendix/front matter takes the volume's median (date_from_volume).
• Cut at 1913-12-31: 6,500 later sections (6.2 M words) of the 1914-1915 volumes are not included.
• Duplicates across the two sources are removed by paragraph id (the website uses the XML's ids). Exact repeats inside the corpus (boilerplate, repeated lists of ministers) are flagged exact_duplicate with duplicate_of.

Flags and is_prose

is_prose is false for rows flagged non_prose (front matter, lists of returns/papers, sections made of short procedural lines), empty, or exact_duplicate. All retained rows are kept so these decisions are visible. Flags can overlap, so their counts should not be added to infer a total number of excluded rows. Filtering on is_prose leaves 211,876 rows; it excludes exact duplicates as well as empty/non-prose records.

flag	rows
web_source	75,885
non_prose	12,355
exact_duplicate	10,606
date_from_volume	3,259
empty	577
date_repaired	423

Size

series	rows	words	prose words
Parliamentary Debates, 1st series (1803-1820)	4,756	14.8 M	14.5 M
New Series (1820-1830)	3,585	13.6 M	13.5 M
3rd series (1830-1891)	93,574	234.0 M	233.7 M
4th series (1892-1908)	80,128	99.0 M	94.3 M
Official Report, House of Commons, 5th series (1909-1913 here)	43,053	40.7 M	40.6 M
Official Report, House of Lords, 5th series (1909-1913 here)	6,217	6.9 M	6.9 M

Series word totals above are rounded to millions; directory totals below are exact. The first and last decades cover only the portion within the release's date range, and rows are partitioned by their retained/repaired date.

directory	shards	rows	words	prose words
data/1800s/	1	1,877	6,709,369	6,529,604
data/1810s/	1	2,863	8,103,112	7,999,188
data/1820s/	1	2,834	11,859,481	11,785,722
data/1830s/	1	10,558	28,840,693	28,820,360
data/1840s/	1	10,552	31,727,336	31,700,707
data/1850s/	1	11,742	33,368,975	33,310,662
data/1860s/	1	16,312	34,420,698	34,332,184
data/1870s/	1	17,466	37,224,693	37,174,505
data/1880s/	1	21,241	59,637,673	59,573,784
data/1890s/	1	35,659	51,980,152	51,616,045
data/1900s/	2	60,950	68,042,989	63,539,781
data/1910s/	1	39,259	37,147,744	37,020,220

Columns

All shards share the same 26-column schema.

column	meaning
doc_id	Record identifier: <zip stem>__<n> (archive) or <volume identifier>__web<n> (website); not a local filename.
date, year	Retained/repaired sitting date as an ISO YYYY-MM-DD string and integer year.
series, volume	Series code as a string (1-4, 5C = Commons 5th, 5L = Lords 5th) and integer volume number.
house	Commons / Lords; null for appendix, front matter and other unassigned records.
title	Section heading as printed.
text	Extracted text, with non-prose elements removed as described above; may be empty.
n_words, n_chars	Integer word and character counts of the retained text.
is_prose	Boolean recommended-training-subset indicator; see flags above.
flags, duplicate_of	List of processing/classification flags and the main record's doc_id for an exact duplicate (otherwise null).
speakers	List of speaker names as printed, in order.
part	debates (one section) / house (sitting text outside a section, or an early whole sitting headed "[TITLE.]") / appendix / other:<tag> (other source elements, including front matter).
col_start, col_end	Nullable integer printed column range.
first_para_id	Nullable Digital Hansard identifier for the first paragraph.
source	archive_xml or historic_hansard_web.
zip	Source archive ZIP stem for XML rows; null for web rows. No source ZIPs are bundled.
source_url	Absolute external source URL: archive ZIP for XML rows or Historic Hansard page for web rows.
web_url	Relative page path under https://api.parliament.uk/historic-hansard/ for web rows (e.g. lords/1805/may/13/roman-catholic-petition); null for XML rows. It is not a local file path.
n_paras, tables, divisions, dropped_words	Integer paragraph, removed-table, removed-division and dropped-word counts.

Verification

verify.json records independent production-run checks of file presence and row counts, unique IDs, ordering, date bounds, markup, flags and duplicate references, plus column coverage against the Historic Hansard column index (1,052,649 of 1,054,331 columns, 99.84 %) and samples re-derived from raw ZIPs and saved pages. Its overall result is PASS (summary.failures is empty). It also records one web re-derivation sample below 0.98 agreement: S3V0242__web00192, approximately 0.889, despite the overall passing status.

Local checks of all 13 released shards confirm their shared schema, every manifest file/row count, unique record IDs, valid duplicate references, nondecreasing dates, date/year/decade consistency, the is_prose rule, and all release word, source, series and flag totals in manifest.json. Directory totals in this README are computed from the released shards. All documented local files and loading paths exist. The source-dependent coverage and raw-text re-derivation checks cannot be independently repeated using this package alone because their source inputs are not bundled.

Known gaps

• Column coverage against the Historic Hansard column index is 99.84 %. The rest is mostly House of Lords sittings of 5-19 June 1893 (4th series vol. 13): the website lists ~190 sections for those days but every one returns 404, and the official archive has no zip for that volume. Smaller shortfalls: 1st series vol. 39 (1819) and a few 4th series volumes, where the website itself lacks some columns.
• 4th series vols 6, 100, 117, 128, 199 and Commons 5th series vol. 33: the archive zip holds only front matter / index and the website volume pages are empty; their date ranges are dissolution or recess periods, so they appear to hold no debates.
• 1st series vols 30-31 have their text but no column numbers in the XML (col_start is null).
• Some speaker names lose a colon in website-sourced rows ("Mr. Manners Sutton Long as..." where the print has "Mr. Manners Sutton: Long as...").

Licence

Contains Parliamentary information licensed under the Open Parliament Licence v3.0.