alea-institute/alea-eu-dates

ALEA EU Dates

Reading and date-reasoning questions over real acts and corrigenda of the Official Journal of the European Union (L series, English, 2001 to 2023). Each row is one training document: an instruction, the act as published (or an excerpt cut at whole elements), one to three questions, and short answers.

The task is to keep apart the dates an EU act carries:

• adoption: the date in the act's title ("of 15 January 2015") and signature;
• publication: the date of the Official Journal issue, shown only in the Official Journal line at the top of the document;
• entry into force: computed from the act's own entry-into-force sentence ("on the twentieth day following that of its publication" = publication + 20 days; "the day following" = + 1; "the day of its publication"; "the third day following"; "the date of its adoption"; or a printed date);
• application: a separate "It shall apply from ..." date, which can be before or after entry into force.

It also covers superseded values: acts that replace one date with another ("'31 July 2017' is replaced by '31 May 2015'": which date applies now, which was replaced), and corrigenda ("for: ... read: ..."), including the difference between the corrected act's original publication date and the corrigendum's own publication date.

Source

All passages come from our KL3M public-record collection of the Official Journal of the European Union (Formex XML, L series, English). EU legal acts are reusable under the Commission's reuse policy (Decision 2011/833/EU). Passages are whole XML elements rendered to text (title, citations, recitals, articles, final provisions, signature, footnotes); omitted parts are marked "[…]". Annexes and tables are never shown. Source text is not edited.

Gold

Gold is computed deterministically from the XML metadata and the act text, never by a language model:

• adoption date from the act's own date element (checked against the date printed in the title);
• publication date from the issue's table-of-contents date (checked against the document metadata);
• entry into force from the single "This Regulation / Decision / Directive ... shall enter into force on ..." sentence. Acts with more than one such sentence, a conditional or split entry into force, or a clause that cannot be parsed exactly are dropped;
• application date only from a single clean "It shall apply from ." sentence. Acts with split application ("except Article 3, which shall apply from ...") or any other apply / take-effect sentence get no application question;
• superseded dates only from acts with exactly one quoted date-for-date replacement;
• corrigenda from the "for:" / "read:" structure and the Official Journal reference in the corrigendum title.

Every question is checked against the rendered passage before a document is kept. The entry-into-force sentence and any application or replacement sentence must be shown. The adoption date must appear in the title. The publication date may appear only in the Official Journal line (or in a printed entry-into-force date that equals it).

Question types

kind	question	answer
open:eif / open:adopt / open:pub / open:apply	on what date did the act enter into force / was it adopted / published / from when does it apply	a date, alone or in one sentence with the computation
mc:<role>	four dates: the other role dates plus off-by-one traps (publication + 19 / + 21 days for twentieth-day acts)	letter (and date)
tf:<role>	"This Regulation entered into force on X." true or false (balanced 50 / 50)	True / False (with the correct date)
role	what a given date marks: adoption, publication, entry into force or application	the role
sup_new / sup_old	which of two dates applies after the act / which was replaced	a date
corr_new / corr_old	which of two wordings the corrigendum establishes / marks as wrong	the wording
orig_pub / corr_pub / orig_adopt	publication date of the corrected act vs of the corrigendum; adoption of the corrected act	a date

Defences against shortcuts:

• Twin documents pair acts of the same year and type with different entry-into-force rules, asked with the same wording.
• Fixed-date acts are included, so "always add 20 days" fails.
• Twentieth-day acts are capped at 30% of acts.
• Multiple-choice positions are uniform.
• True / false labels are balanced.
• The asked role is balanced per act type.
• Some documents start with one or two worked examples from other acts, which never share a date with the target.

Formats: plain text, question-answer frames and chat turns, roughly 40 / 30 / 30.

Columns

id, text (the full training document), family (eu_dates), kind (first question's kind), source (eu_oj), gold (JSON list: one object per question with the kind, the question, the gold value or label, options where relevant, and the answer text), presentation (plain / qa / chat), group_id (twin documents and documents built from the same act share it), year, act_type (Formex legal value, e.g. REG, REG_IMPL, DEC, DIR, CORRIGENDUM), eif_rule.

Counts and composition

• 41,685 documents, about 23.6M tokens (alea-v2b-sw32k tokenizer); 63,283 questions (1 question in 60% of documents, 2 in 28%, 3 in 12%).
• Built from 25,218 acts (of 48,495 main acts scanned, 32,706 had a cleanly parseable entry-into-force sentence; near-identical daily acts were capped) and 2,061 corrigenda; publication years 2001 to 2023 (all but 22 documents from 2004 onwards).
• Entry-into-force rule (documents): twentieth day after publication 30.1%, day following 20.0%, printed date 12.4%, date of adoption 11.9%, day of publication 10.5%, third day 7.1%, other Nth day 3.0%; corrigenda 4.9%.
• Act type: Regulation 38.7%, Implementing Regulation 25.8%, Decision 16.8%, corrigenda 4.9%, Delegated Regulation 4.7%, Directive 3.9%, Implementing Decision 3.7%, other 1.5%.
• Authoring institution: Commission 69.6%, Council 23.0% (two metadata codes), European Parliament and Council 4.1%, European Central Bank 0.8%, others 2.5%.
• Years 2004 to 2023 each 3.9% to 5.9% of documents; no year x act type stratum above 5.0%.
• First question of each document: entry into force 75.7% (open 34.0%, true/false 22.1%, multiple choice 19.6%); the rest ask adoption, publication, role, application, superseded dates and corrigenda (further questions in a document mostly ask these other dates).
• 84.9% of act documents belong to twin pairs (two acts of the same year and type with different entry-into-force rules).
• 8.1% of documents start with one or two worked examples.
• Presentation: plain 39.7%, question-answer 30.4%, chat 29.9%.

Checks

Surface-cue baselines on the released rows (5-fold cross-validated bag-of-words models that see only the question):

check	result	prior
true / false, question only	0.508	0.504 (majority)
true / false, question only, dates masked	0.511	0.504
multiple choice, always the most frequent letter	0.252	0.25
multiple choice, question stem only	0.261	0.25
two-option questions, gold is the first-listed option	0.492	0.5
role question, question only, dates masked	0.323	0.332 (majority)
entry into force, always publication + 20 days	0.307	(share of twentieth-day acts)
entry into force, always the publication date	0.213
entry into force, the latest date printed in the passage	0.175

A role-question model that sees the raw dates reaches 0.414 because many acts in one Official Journal issue share their publication and adoption dates; with folds grouped by date it reaches 0.395.

Decontamination: rows sharing long word n-grams with held-out probe items, evaluation benchmarks or other held-out sets were removed (1,371 rows; most are near-identical sibling acts of held-out acts), and 71 rows were removed to keep every year x act type stratum at or below 5%. About 1.5% of act groups are held out and not released.

Review

An independent reviewer saw 60 randomly drawn rows (stratified over all question types) as rendered text only, without the gold labels, and redid every date computation by hand. Result: 57 correct, 3 correct with minor wording issues, 0 wrong. The minor issues were "removes" wording for a replaced date that also survives in an inserted paragraph, an "Answer with a date" hint on a question whose answer is an event, and a clumsy explanation phrase on a printed-date answer. Before that, the builders read 35 rows. That read found one wrong answer type (a signature date used as an adoption date in a corrigendum), which was fixed in the generator before release.

In multiple-choice entry-into-force questions the wrong options are usually the publication date, the adoption date and a date one day off the correct one. The answer still has to be computed from the rule.

Limits

• The authoring institution is dominated by the Commission. Rows are therefore capped by year x act type (no stratum above 5%), not by author. Near-identical daily acts (for example standard import values) are capped at 40 acts per title pattern.
• Only articles needed for the questions are guaranteed to be shown; long acts are excerpts.
• Acts that enter into force on notification or under a condition (for example "provided that all the notifications ... have been made") are dropped, not modelled.
• English only; the same acts exist in other EU languages but are not used here.