biglam/old_bailey_proceedings

Old Bailey Proceedings, 1674–1913

The published accounts of 197,751 criminal trials held at the Old Bailey (from 1834 the Central Criminal Court), London, between April 1674 and April 1913, plus 475 Ordinary's Accounts — pamphlets by the chaplain of Newgate on prisoners executed at Tyburn, 1676–1772. Text and structured markup were created by the Old Bailey Online project (University of Sheffield, University of Hertfordshire, Open University; directors Clive Emsley, Tim Hitchcock and Robert Shoemaker, project manager Sharon Howard). This dataset is a repackaging of the project's XML data, version 7.2 (March 2015), published at 10.15131/shef.data.4775434.v2: one row per trial, with the project's tagging of people, offences, verdicts and sentences turned into columns, and the TEI kept verbatim.

Roughly 123 million words of trial text. The Proceedings are the largest body of texts about the lives of non-elite people ever published; they are also a published source — edited, selective and shaped by the printers' commercial interests, on which see the project's own publishing history.

Configs

config	rows	what
trials (default)	197,751	one row per trial account: text, verbatim TEI, typed persons, offences, verdicts, punishments and the joins between them
sessions	2,163	one row per printed session (the source file): full text incl. front matter and summaries, the XML file, all persons, and summary_items
ordinarys_accounts	475	one row per Ordinary's Account pamphlet: text, XML, persons

from datasets import load_dataset

# trials — ~550 MB, one row per trial; stream unless you want the `xml` column locally
ds = load_dataset("biglam/old_bailey_proceedings", "trials", split="train", streaming=True)

# Ordinary's Accounts are small
oa = load_dataset("biglam/old_bailey_proceedings", "ordinarys_accounts", split="train")

DuckDB reads the parquet directly and pushes column projection down, which matters here because xml is half of every file:

SELECT date, offences[1].category, verdicts[1].category
FROM 'hf://datasets/biglam/old_bailey_proceedings/trials/*.parquet'
WHERE n_defendants > 1;

trials — columns

column	type	description
trial_id	string	Old Bailey Online id, e.g. t18381217-265; obo_url opens the trial on the project site
session_id, date	string	session file id and its opening date, YYYYMMDD (kept as a string because it is the id)
part	int	1 = April 1674–October 1834, 2 = November 1834–April 1913 (the two digitisation phases, see Provenance)
text	string	the trial text, whitespace-normalised, markup removed
xml	string	the trial's <div1 type="trialAccount"> verbatim — character offsets for every tagged span, and anything this schema does not model
persons	list of struct	every <persName> in the trial: id, type, name, given, surname, gender, age, occupation, place, aliases
offences	list of struct	id, category (9), subcategory (56), text, crime_date, place, victim_ids
verdicts	list of struct	id, category (4), subcategory (24), text
punishments	list of struct	id, category (6), subcategory (27), text, defendant_ids, found_in
charges	list of struct	the project's criminalCharge links: defendant_ids → offence_ids → verdict_ids, plus unresolved_ids
n_defendants, places		count of defendants; <placeName> strings in the trial

The category vocabularies are the project's own — offence categories theft, deception, royalOffences, violentTheft, breakingPeace, sexual, kill, miscellaneous, damage; verdicts guilty, notGuilty, miscVerdict, specialVerdict; punishments imprison, transport, miscPunish, death, corporal, noPunish. Their meanings are explained on the project's crime, verdict and punishment pages and should be read there before use; theft alone is 72% of offences.

	count
offences	211,111 (theft 150,992 · deception 15,869 · royalOffences 10,680 · violentTheft 9,012 · breakingPeace 7,686 · sexual 7,219 · kill 5,375 · miscellaneous 3,224 · damage 1,054)
verdicts	224,076 (guilty 164,068 · notGuilty 59,371 · miscVerdict 491 · specialVerdict 146)
punishments	168,266 (imprison 88,588 · transport 41,204 · miscPunish 12,932 · death 11,600 · corporal 9,024 · noPunish 4,918)
charges	257,057
trials with more than one defendant	18.2%

Persons

1,101,978 persName elements: 312,366 witnesses, 246,095 defendants, 203,041 victims, 1,276 jurors, 310 judiciary — and 338,890 with type null. The null is the source, not a parsing gap: in part 1, names were tagged automatically (GATE) and only defendants and victims were then typed by hand, so 58.7% of part-1 person tags carry no role; in part 2 it is 0.5%.

Person attributes are sparse because the Proceedings are. Across all persons: gender 99.5%, age 13.1%, occupation 7.6%, place 1.3%. For defendants: age 58.1%, occupation 11.0%, place 5.8%, alias 2.6%. A null means the text does not state it. Defendant gender: 192,427 male, 53,404 female, 264 indeterminate.

name is the tag's own text (nested names excluded, so "George Palfreman and Mary his Wife" yields Mary Palfreman for Mary); 414 persName elements are empty in the source and have name null. Where the project recorded two age values for one person (298 cases) the first is kept — the second is in xml.

Joins, and where the source is ambiguous

charges reproduce the project's criminalCharge joins, which is how a multi-defendant, multi-offence trial says who was convicted of what. 1,455 charges name a target id that does not exist anywhere in the file (unresolved_ids, e.g. t18381217-378-offence-3); they are kept rather than dropped.

Reused ids. In 916 trials a verdict id, and in 90 an offence id, is used twice inside the same trial in the source (typically a GUILTY and a NOT GUILTY both tagged …-verdict-1). Both elements are present in verdicts/offences, but a charges row pointing at that id is ambiguous. Resolve by position or from xml.

Punishments are attached through the project's defendantPunishment joins. Before November 1834 sentences were often printed in a separate summary at the end of the session, so a punishment's text and its defendant can live in different places: found_in names the trial div that holds the sentence text (514 rows where that is a different trial), and the same punishment then appears in both trials. 106 punishments have no join at all (defendant_ids empty). 3,324 guilty verdicts have no linked punishment in the source.

sessions

One row per session file: text (everything, 124 million words — front matter, juror lists, advertisements, summaries, trials), xml (the whole file, verbatim, in its declared encoding), persons (all 1.2 million, typed as above), n_trials, div_types, and summary_items: the offences, verdicts, punishments, aliases and occupations that the earliest sessions record in supplementaryMaterial and punishmentSummary divs rather than in trial accounts — 1,504 punishments, 105 verdicts, 101 offences, each with its category, text, containing div and joined ids. Three sessions have no trial accounts at all: 16811124 and 16830524A are front matter and summary only; 16781211 is 74,000 characters of text with no markup whatsoever in the source.

ordinarys_accounts

475 pamphlets, 17 May 1676 – 14 October 1772, 2.6 million words, 21,827 tagged persons. No offence/verdict/punishment markup exists for them in the source.

Coverage

decade	trials	decade	trials
1670s	491	1800s	7,760
1680s	2,510	1810s	12,260
1690s	3,142	1820s	18,152
1700s	812	1830s	22,711
1710s	2,999	1840s	25,574
1720s	4,916	1850s	12,393
1730s	4,672	1860s	10,147
1740s	4,222	1870s	8,002
1750s	4,057	1880s	9,924
1760s	4,206	1890s	8,399
1770s	5,993	1900s	7,889
1780s	7,892	1910s	2,656
1790s	5,972

The dip after 1850 is the source, not the digitisation: from the 1850s the Proceedings reported fewer trials at length and eventually only those the editors chose. Coverage of the 1670s–1700s is thin because few early Proceedings survive.

Provenance

The project describes its methods in full at oldbaileyonline.org/about; what follows is the short version.

• Text. Part 1 (1674–October 1834) was transcribed from microfilm of the original printed Proceedings by double rekeying — typed twice by two typists and reconciled — which is why this corpus has far fewer transcription errors than OCR-derived historical text. Part 2 (November 1834–1913) went through an automated markup pipeline at the Digital Humanities Institute, Sheffield, with checking by a team at the University of Hertfordshire. Page images were scanned at 400 dpi; they are not included here (<xptr> references to them survive in xml).
• Markup. Part-1 markup was done by hand by a team of data developers at the Humanities Research Institute, Sheffield, except person names, which were tagged automatically with GATE (~80–90% of names caught; forename-and-surname names only; the untyped-name gap above). Part 2 was tagged automatically and checked by hand.
• This repackaging. scripts/parse_obo.py (lxml + polars) reads the 2,163 session and 475 Ordinary's Account files and emits these three configs; it is reproducible from the figshare zip. Two independent code reviews and an invariant check of 3,094 trial divs against the raw XML were run before release. Nothing in the source was corrected: reused ids, empty names, unresolved joins and the unmarked session are preserved as found and recorded above.
• Previous version. The 2022 Hub version of this dataset (one row per session, from the BigLAM datasets sprint, contributed by @shamikbose) remains at revision e96ea6d. Its card stated CC BY 4.0; that was an error in the loading script.

Licence

CC BY-NC 4.0, the licence the Old Bailey Online project attaches to all its data (licence.txt in the source archive). The project's own guidance: its interpretation of "commercial" normally excludes academic research outputs even when published with commercial publishers; permissions beyond the licence may be available from the project. Reuse should include the statement the project asks for:

This work builds on Old Bailey Online that is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License. Permissions beyond the scope of this license may be available at https://www.oldbaileyonline.org/about/legal-info.

Citation

@misc{howard2017oldbaileyxml,
  author = {Sharon Howard},
  title  = {Old Bailey Online XML Data},
  year   = {2017},
  doi    = {10.15131/shef.data.4775434.v2},
  url    = {https://doi.org/10.15131/shef.data.4775434.v2}
}

Cite individual trials by id (t18381217-265) and the project name and URL, per the project's citation guidance.

Repackaged by Daniel van Strien (Machine Learning Librarian, Hugging Face); all source data and markup created by the Old Bailey Online project. Maintenance: limited — revised if the project publishes a new XML version.