alea-institute/alea-heading-gen

ALEA Heading and Title Generation

Write, pick or match the real heading of a legal or government text. Each example shows a verbatim passage from a public-record document with its heading, title or caption removed. The answer is the real heading from the source document, copied exactly.

The data was built for mid-training of ALEA language models. Every example is a complete training document: an instruction or user turn, the passage, and a short answer.

Tasks

kind	input	answer	docs
bill_title	a congressional bill's text, with its front matter and title cut	the bill's official (long) title, e.g. "To amend ... , and for other purposes."	2,550
usc_heading	a U.S. Code section body, with its heading line cut	the section catchline	1,288
heading_mc	a U.S. Code or CFR section body	its heading among 4 headings from the same chapter or part	1,990
ecfr_heading	a Code of Federal Regulations section body, with its heading line cut	the section heading	868
case_name	a court opinion, starting at the author line, with the caption, docket and counsel lines cut	the case name ("A v. B")	1,381
fr_title	a Federal Register document's SUMMARY and DATES block	the document title	448
heading_match	3–4 short sections from one chapter or part	the pairing of each section with its shuffled heading	135

How the gold was made and checked

• Passages are verbatim, contiguous cuts: whole paragraphs, or whole lines for bills and opinions. Text is never edited, and the heading or title is never shown.
• The answer must be supported by the text:
  • Sections under 40 words, "[Reserved]", repealed and similar sections are skipped.
  • FR and bill titles need at least half of their content words in the shown text.
  • A case is kept only if both parties' names appear in the opinion text. No shown line may name both parties around "v.".
• Every detail of the gold must be in the shown text. A title or heading is dropped if it contains a number or year, a name or capitalised word, or (for section headings) a content word of 5+ letters that does not occur in the passage. Generic words such as Act, Rule or Program are tolerated. Passages that begin or end with a neighbouring section or block heading are dropped.
• No copying. Section headings of 3 or more content words that appear verbatim in the body are dropped. A case name or title that appears verbatim in the shown text is dropped too.
• Multiple choice and matching. Distractor headings come from different sections of the same chapter or part, under four rules:
  • each must differ in wording from the gold;
  • its own section text must not be similar to the target text;
  • its words must not mostly appear in the target passage;
  • it must not be generic. "Definitions", "Scope.", "General." and similar headings are never options, and are never golds in multiple choice.
• Repetition is capped. The same heading can be the answer at most 12 times per kind, and the same FR title at most 3 times.

Answer formats and presentation

Formats:

• Generation answers: the bare heading (54% of all examples), a short frame such as "The heading is: ...", or a JSON object when the instruction asks for one and names its key.
• Multiple choice: the letter, the letter and the heading, the heading alone, or JSON.
• Matching: pairs ("1-B, 2-A"), one pair per line, or a JSON map.

Each kind has 8–9 instruction wordings. Documents are rendered as plain text (40%), question / answer frames (30%) and chat turns (30%). About 1.8% of examples include one or two worked examples from other documents with different answers.

Composition (8,660 documents, 5.52M tokens)

source	docs	share
U.S. Code sections	2,598	30.0%
congressional bills	2,550	29.5%
Code of Federal Regulations	1,683	19.4%
court opinions (Caselaw Access Project)	1,381	15.9%
Federal Register	448	5.2%

No source exceeds 30% of documents. No single title, agency, court or congress-and-bill-type exceeds 5%; the largest are 26 CFR, 46 CFR and 42 U.S.C. at 5.0% each. There are 396 distinct strata. The median document is 755 tokens and the longest is 1,171.

Shortcut checks

• Multiple choice:
  • option text alone: 0.21 against a chance prior of 0.25;
  • longest option: 0.25; shortest option: 0.25;
  • most frequent gold letter: 0.27.
• Matching: an identity-mapping baseline scores 0.30 against a prior of 0.31.
• Generation: the answer appears verbatim in the prompt in:
  • 0% of FR titles;
  • 0.1% of bill titles and case names;
  • 17–20% of U.S. Code / CFR headings, all one- or two-word headings such as "Reports" or "Pacific Area.".

Sources and filtering

All passages come from public-record documents in the KL3M corpus:

• the U.S. Code;
• the eCFR;
• the Federal Register;
• congressional bill texts;
• court opinions from the Caselaw Access Project.

Party, judge, official and agency names are left as published.

Filtering:

• Passages containing e-mail addresses or phone numbers are dropped.
• Every example was screened against evaluation benchmarks and held-out probe sets, at the 8-gram and 13-gram level. Matching examples were removed.
• About 1.5% of source documents are held out and are not in this release.

Review

The builder read 35 released examples against their source metadata and found none wrong. An independent blind review of 60 examples found 3 wrong items, all titles carrying details absent from the shown text. The detail-support filter above was then added, and a second blind review of 60 new examples found 1 wrong item (a bill title naming a part of the bill cut off from the excerpt). Defects found in earlier rounds were fixed before release by dropping rows or fixing the generator: a caption leak in case names, and ambiguous generic headings in multiple choice.

Generation answers are the real headings. Other wordings of a heading can also be reasonable.

License

CC BY 4.0. The underlying documents are U.S. government works and public court records.