Hansard - UK Parliamentary Debates, 1803-1913

UK parliamentary debate sections and related sitting, appendix and front-matter records drawn from Parliament's Digital Hansard XML archive and Historic Hansard website.

The dataset contains 231,313 records and 409,086,680 whitespace-separated words. The recommended training subset, is_prose = true, contains 211,897 records and 403,427,256 words.

Parquet is the authoritative deliverable: 13 shards in 12 decade directories, using a shared 26-column schema. A faithful JSONL export is also included. No dataset-loading script is required.

The earliest and latest stored record dates are 1803-03-29 and 1913-12-12. The collection cutoff is 1913-12-31, not the date of the final sitting represented. The collection has known source gaps; it should not be described as every parliamentary contribution without exception.

Files and direct size counts

Parquet files total 819,978,715 bytes. manifest.json provides exact per-file sizes, row counts and SHA-256 hashes.

Directory	Shards	Records	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	31,727,537	31,700,908
data/1850s/	1	11,742	33,368,975	33,310,662
data/1860s/	1	16,312	34,421,087	34,332,573
data/1870s/	1	17,466	37,225,685	37,175,497
data/1880s/	1	21,241	59,641,124	59,577,235
data/1890s/	1	35,659	51,990,213	51,625,619
data/1900s/	2	60,950	68,050,638	63,547,645
data/1910s/	1	39,259	37,148,766	37,022,243
Total	13	231,313	409,086,680	403,427,256

Every directory contains part-0000.parquet; data/1900s/ additionally contains part-0001.parquet. Directory assignment uses the retained/repaired date. The first and last decades cover only the collection's date range.

Record text totals 2,342,378,014 Unicode characters and 2,346,360,520 UTF-8 bytes, excluding metadata and serialization overhead. Words use len(text.split()); characters use len(text). These are direct integrity counts, not external tokenizer or pipeline-statistics results.

JSONL export

data/export.jsonl is generated from the final written Parquet and contains all 231,313 records and all 26 columns, including source provenance. There is no additional text transformation or filtering. JSON objects follow sorted shard order and the row order within each shard.

• UTF-8, one JSON object per line.
• Size: 2,525,383,257 bytes.
• SHA-256: 59d98a4039a70e7b2568477b428b5ca50068cc30c0ab44ad8af6db37b13deeea.

Supporting files:

• manifest.json: full file inventory, measured totals, flags and series/decade counts.
• verify.json: independent complete Parquet/JSONL validation.

Original XML ZIPs, cached web pages, printed scans and raw model responses are not bundled. Source fields identify external resources, not files that should exist under data/.

Coverage by Hansard series

Series	Period represented	Records	Words	Prose words
1: Parliamentary Debates, first series	1803-1820	4,756	14,832,317	14,548,628
2: New Series	1820-1830	3,585	13,567,201	13,492,519
3: third series	1830-1891	93,574	234,021,422	233,698,813
4: fourth series	1892-1908	80,128	99,061,373	94,266,138
5C: Commons Official Report, fifth series	1909-1913	43,053	40,663,025	40,564,251
5L: Lords Official Report, fifth series	1909-1913	6,217	6,941,342	6,856,907

Schema - 26 columns

Column(s)	Type	Meaning
doc_id	string	Unique record id: archive ZIP stem plus ordinal, or volume identifier plus web ordinal. Not a local filename.
date, year	string, int64	Retained/repaired ISO sitting date and corresponding year. Some front matter has an inferred volume date.
series, volume	string, int64	Series code (1-4, 5C, 5L) and volume number.
house	nullable string	Commons or Lords; null for some unassigned/front-matter records.
title	nullable string	Section heading as supplied by the source.
text	string	Extracted section/sitting text with paragraph breaks. May be empty.
n_words, n_chars	int64	Whitespace-separated word count and Unicode character count of stored text.
is_prose	bool	Recommended training-subset indicator.
flags	list of string	Source, classification, date and duplicate flags described below.
duplicate_of	nullable string	Canonical retained record's id for a flagged duplicate.
speakers	list of string	Speech-speaker names in source order; not every personal name appearing in the text.
part	string	debates, house, appendix or other:<tag>.
col_start, col_end	nullable int64	Printed column range, when available. Not page numbers or a proof of gap-free text.
first_para_id	nullable string	First available source paragraph identifier. Restored list items do not necessarily have separate source paragraph ids.
source	string	archive_xml or historic_hansard_web.
zip	nullable string	Archive XML source ZIP stem; null for web-sourced rows.
source_url	string	Absolute source ZIP or Historic Hansard page URL.
web_url	nullable string	Relative Historic Hansard page path, under https://api.parliament.uk/historic-hansard/.
n_paras	int64	Retained source paragraph/unit count, including restored list/block units. Nested headings and composite units mean this need not equal the number of blank-line-separated display paragraphs.
tables, divisions, dropped_words	int64	Removed-table, removed-division and intentionally dropped-word counters. These do not count recovered extraction omissions.

Debate rows represent top-level sections. Nested sub-section titles can remain in their text. house rows collect sitting-level material outside individual debate sections, or an early whole sitting where the XML lacks section structure. Oral questions and written answers are usually separate debate records.

Text assembly and training mask

Primary source: the official Digital Hansard XML archive. The Historic Hansard website supplements missing/incomplete archives and individual missing columns. 75,885 records, containing 64,064,164 words, have web provenance.

Speeches retain source speaker labels and historic wording. Paragraphs are separated by blank lines. Tables, division-name lists, indexes and other non-prose components are excluded from training text according to the extraction policy. Ancillary/front-matter records remain stored and flagged. Corpus-informed joining repairs some words broken across column/page boundaries while retaining common historical hyphenation.

Dates contain source/OCR mistakes. The date-processing policy uses nearby valid dates or a volume median for outliers/front matter, with explicit flags. These repaired/inferred dates are approximate, not independently verified dates for every source page.

is_prose is false if any of non_prose, empty or exact_duplicate is present. Flags can overlap; their counts should not be summed to infer the excluded-row total.

Flag	Records
web_source	75,885
non_prose	12,336
exact_duplicate	10,592
date_from_volume	3,259
empty	577
date_repaired	423

The exact_duplicate flag uses canonical matching, not guaranteed byte-for-byte equality: letter case, punctuation and spacing can differ. Canonical matching preserves letters and digits but removes punctuation, including potentially meaningful signs. All flagged records remain available, so users requiring literal repetitions or punctuation-sensitive comparisons should inspect text rather than treating the flag as a semantic-equivalence guarantee.

Source-backed restoration and review

Source review examined 7,855 selected Historic Hansard pages-risk cases and stratified control samples-and all 544 referenced XML source ZIPs. The XML checks examined 3,634 narrative/list subtrees; their normalized text was checked against the released records, including nested-list and interleaving cases. These checks did not establish scan-level OCR accuracy.

178 source-backed text blocks in 138 records, containing 23,765 source words, are incorporated. They include legislative clauses, financial resolutions, quotations and factual enumerations represented by irregular HTML list/quotation markup. Source-localized insertions retain the existing text character-for-character rather than rewriting surrounding speeches. Substantial legislative text is classified as prose where appropriate, and affected duplicate references are made consistent with the stored text.

The six numbered resolutions in the 8 August 1878 “Navy and Army Expenditure, 1876-7” debate are present, including the financial amounts as supplied by Parliament's digital source. Source numerical inconsistencies and OCR slips are not silently corrected by inference.

Verification and limitations

Independent validation reads every written Parquet record and every JSONL object. It checks:

• shared field schemas, unique ids and record counts;
• date validity, chronological order and decade partitioning;
• provenance, prose-mask and duplicate-reference consistency;
• stored word/character lengths and nonnegative counters;
• printed-column bounds;
• complete Parquet/JSONL equality, file sizes and SHA-256 hashes;
• manifest totals, flags and series/decade counts;
• presence of all six financial resolutions in the identified debate.

Structural validation is not proof of perfect transcription or exhaustive coverage. The text is Parliament's digitized historical edition, with inherited OCR errors and editorial/transcription inconsistencies. Occasional corrupted words, speaker-label punctuation loss, remaining extraction omissions and inaccurate/inferred dates can remain. Review is targeted, not manual proofreading of 409 million words. The website and XML often represent the same underlying OCR and are not independent printed-page witnesses.

Known source limitations include:

• The source assembly's column-index coverage report gives 1,052,649 of 1,054,331 columns (99.84%). This historical coverage figure was not independently remeasured against a fresh complete column index.
• Particularly incomplete: Lords sittings of 5-19 June 1893, fourth-series volume 13; the assembly notes report inaccessible section pages and no usable archive ZIP. Smaller column shortfalls affect first-series volume 39 and several fourth-series volumes.
• Fourth-series volumes 6, 100, 117, 128 and 199, and Commons fifth-series volume 33, have unusable/front-matter-only source coverage in the assembly notes. Some date ranges correspond to recess/dissolution periods.
• First-series volumes 30-31 lack printed column numbers in their XML.
• Website-sourced speaker labels can lose a colon, e.g. a name running into the following sentence.

For local validation:

pip install -r requirements.txt
python tools/verify.py
python -m unittest discover -s tests -v

Licence and attribution

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

Please credit UK Parliament / Historic Hansard as the source. See the licence for its attribution requirements and exclusions. The source URLs in each record support provenance and further source checking.