BangorAI/deddfwriaeth-cymru-1

Deddfwriaeth Cymru 1: Welsh legislation (cy–en)

Welsh–English pairs of legislative provisions from legislation.gov.uk, covering everything made in Wales with both language versions published there:

type	type	items	provision pairs
Acts of Senedd Cymru	asc	34	25,840
Acts of the National Assembly for Wales	anaw	44	28,737
Measures of the National Assembly for Wales	mwa	22	7,159
Wales Statutory Instruments	wsi	4,496	212,203
total		4,596	273,939

Years 1999–2026. The text is each item as originally enacted or made, not later revised versions. Retrieved on 5 and 6 November 2026.

How it was made

1. Pairing is exact, not statistical. The English and Welsh XML (Crown Legislation Markup Language) give every provision the same element id, such as section-1-a or schedule-2-paragraph-3. Each provision's own text is paired with the provision of the same id in the other language. Nested sub-provisions are separate rows, so a row is often a clause fragment ending in "—" or ";", exactly as it appears in the legislation.
2. Excluded. Annotations, commentaries, footnotes, explanatory notes and signatures. Also excluded are items whose "Welsh" file is actually the English text, which legislation.gov.uk serves as a fallback.
3. Length check. suspicious_length marks pairs whose length ratio is unusual for cy–en (2,074 rows).
4. De-duplication. Pairs are compared after lower-casing, collapsing whitespace and replacing every run of digits with 0. Repeated standard wording, for example commencement and citation provisions that differ only in a year or number, is kept once, with n_occurrences recording how often the pattern appeared. Provisions with no words on the Welsh side are removed.
5. One-to-many. one_to_many marks a Welsh or English string that is paired with five or more different translations (2,473 rows).

Recommended filter for MT training data (267,105 rows, about 9.6 million Welsh and 9.0 million English words):

from datasets import load_dataset
ds = load_dataset("BangorAI/deddfwriaeth-cymru-1", split="train")
clean = ds.filter(lambda r: not r["suspicious_length"] and not r["one_to_many"] and r["cy"] != r["en"])

Fields (default config)

field	description
cy, en	provision text
provision_id, seg_index, element_kinds, pnumber_en, pnumber_cy	provision id, position, XML element kinds and the provision number on each side
length_ratio, suspicious_length, one_to_many	quality checks
n_occurrences, pair_hash	de-duplication count; sha1 of cy\ten
legislation_id, type, year, title_en, title_cy, url_en, url_cy	the legislation item
source_id, bucket, licence, licence_url, attribution	provenance on every row

The documents config has one row per legislation item with a Welsh version (4,610). Each row gives the titles, URLs, the number of paired provisions before de-duplication (n_segments) and the number of provisions found only in English or only in Welsh. Fourteen short items have no rows in the default config, because all their provisions were repeats of standard wording. Another 2,105 Wales Statutory Instruments were made in English only (mostly local traffic orders) and are not included.

Source and licence

Source: legislation.gov.uk, managed by The National Archives. The site's terms say: "All content is available under the Open Government Licence v3.0 except where otherwise stated." This dataset includes legislation made in Wales only. It does not include EU-derived legislation, which has different terms.

Contains public sector information licensed under the Open Government Licence v3.0. Mae'n cynnwys gwybodaeth sector cyhoeddus a drwyddedir o dan y Drwydded Llywodraeth Agored v3.0.