Datasets:

---

biglam

/

old_bailey_proceedings

like 4

Follow

BigLAM: BigScience Libraries, Archives and Museums 206

Ready for review. Final state of this PR after two review rounds of the parser.

What's in it

• trials (default) — 197,751 rows, one per <div1 type="trialAccount">: text, verbatim TEI xml, typed persons (defendant / victim / witness / juror / judiciary; type is null where the source records no role — 338,892 part-1 names) with given / surname / gender / age / occupation / place / aliases; offences (9 categories / 57 subcategories, crime date, place, linked victim ids); verdicts; punishments (category / subcategory, defendant_ids, found_in = the trial div holding the sentence text); charges (the criminalCharge joins: defendant → offence → verdict, plus unresolved_ids for targets that resolve to nothing); places; obo_url.
• sessions — 2,163 rows: full session text + file xml, typed persons, div_types, and summary_items (offences / verdicts / punishments / aliases recorded in supplementaryMaterial / punishmentSummary divs rather than trial accounts — the 1670s–80s style; 1,504 punishments, 5,049 such defendants).
• ordinarys_accounts — 475 rows, same person schema.
• scripts/parse_obo.py — the parser (lxml + polars), reproducible from the figshare zip.
• Licence corrected to CC BY-NC 4.0 (source licence.txt; the 2022 loading script hard-coded CC BY 4.0). dataset_info dropped (schema changed). Previous session-level parquet remains at revision e96ea6d.

Verification

• Two independent code reviews of the parser (12 + 10 findings; all bugs fixed, the rest confirmed ok with evidence).
• Invariants check of 3,094 trial divs against the raw XML: every person id / interp, offence & verdict id and category, charge partition, punishment link and in-div punishment matches.
• All three configs read from this branch with DuckDB (hf://datasets/biglam/old_bailey_proceedings@refs%2Fpr%2F4/<config>/train-*.parquet).

Source quirks preserved, not hidden (to be documented in the card)

• 916 verdict ids and 90 offence ids are reused within a trial div in the source → charges.verdict_ids / offence_ids are ambiguous there.
• 298 persons carry two age interps (first kept); 414 <persName> elements are empty (name null).
• 3 sessions have no trial divs (16811124, 16830524A = front matter/summary only; 16781211 has 74k chars of text and no markup at all) — present in sessions only.
• 1 duplicate trial id in the source (s16980504-1 → second copy s16980504-1-dup2); 106 punishments with no join; 514 punishments whose text sits in a different trial div than their defendant (found_in).
• Sparse-by-source: occupation 7.6% of persons, person place 1.3%, offence place 28%, crime date 56%, age 13% of persons (58% of defendants).

Not in this PR

• Card rewrite (curator voice, DHI attribution statement, sparsity table, changelog crediting the 2022 sprint) — follow-up PR.
• Note: the Hub viewer only renders main, so schema/xml-cell rendering can only be eyeballed after merge.