alea-institute/alea-schema-extract

ALEA schema extraction

Schema extraction from real public-record documents. Each row asks for a structured record of ONE document in a target schema: a JSON object (flat or nested, with row-specific key names), a field | value table, a fill-in template, a JSON array of records with a filter (parties of one role, Items of one section, exhibits of one series), or a conversion where a worked example in one schema and format is shown for another document and the target uses different keys and a different format. Fields that the document does not state must follow the null convention stated in the instruction (null, "not stated", omit the key, N/A, "none given" or "unknown").

No language model wrote any part of this data. Instructions are hand-written templates (10+ wordings per kind); every gold value comes from code over the document text and its metadata; worked explanations are rendered by code from the evidence line of each field.

Sources (public record)

source	documents shown	gold from
Federal Register rule and proposed-rule preambles (daily-issue XML, 2000-2023)	agency heading through DATES paragraph, as printed	XML elements (agency, sub-agency, CFR lines, docket line, RIN, subject, ACTION); effective date / comment deadline only from plain DATES sentences, any other date-bearing sentence makes the field unaskable
CourtListener district-court civil dockets	PACER-style docket sheet: a header in one of four real label layouts, parties with roles (only when every party has one clean role), the first 6-14 numbered entries verbatim	docket metadata (judge, referral judge, cause, nature of suit, jury demand, filed / terminated dates); entry-derived fields by fixed rules over the shown entries
Form 8-K current reports (EDGAR)	the body from the first Item head through the signature block; the cover page is never shown	strict parser: Item heads, exhibit lists, signature block; any ambiguity drops the field or the document

Composition of the released rows: 5,368 documents, 4.16M tokens (alea-v2b-sw32k); court dockets 40%, Federal Register 35%, Form 8-K 25%; 424 strata (agency / court / filer), none above 3% of rows. Kinds: JSON 38%, table 20%, template 20%, record lists 16%, conversion 5%. 35% of rows show a short worked explanation before the answer; each has a direct-answer twin of the same item in another row of its group. Presentations: plain 40%, question/answer frame 29%, chat 31%.

Gold and abstention

• 17% of requested fields are not stated in their document; every such null is verified absent (the label or sentence is not printed in the shown text). Groups pair two documents asked with the same schema, and the schema always holds a field that is stated in one document and not in the other, so the question alone does not reveal the null.
• A question-only classifier (bag of 1-2-grams, grouped 5-fold) predicts null vs stated no better than the per-field prior on every field with enough rows (largest lift +1.6 points).
• Every answer is parsed back with a real parser (JSON / table / key: value) and equals the stored gold object, key order included. In conversion rows no value of the example answer appears in the target answer.
• Rows with an e-mail address, phone number, street address (with or without house number), P.O. box or postcode were dropped, not edited. Public-record names (parties, judges, companies, officials) are kept.

Review

The builder read 204 rows across six builds (30 on the final build, no errors) against their source passages and fixed every generator issue found (8-K heading and "Not Applicable" Item handling, entry-rule edge cases, borderline sub-agency nulls). A blind review of 60 fresh rows by an independent reviewer that saw only the rendered rows found no wrong values, wrong nulls, schema violations or copied examples (4 minor notes on wording and source-text artifacts).

Release v2 (added rows)

v2 adds 6,118 rows (5.08M tokens) to the 5,368 rows of release v1; v1 rows are unchanged.

Contract excerpts (3,416 rows). Excerpts of EDGAR exhibit-10 agreements. The main excerpt is the opening of the agreement (title, opening paragraph naming the parties, recitals, and the term clause when it is near the start); for short agreements it can run on, in one piece, through the governing-law clause. It is sometimes followed by a second verbatim excerpt from later in the same agreement (a numbered clause, the governing-law clause or the signature page), labelled only by its section number ("Excerpt 2 (Section 14)") or "later in the agreement". Each row asks for only 2-4 fields (parties, the labels the opening paragraph gives them, the agreement date, the effective date, the title, the initial term, sometimes the governing law), and the answer contains exactly the requested keys:

• 51% of rows: fields people often expect in a contract record (signatures, fees, governing law, notice details) are not in the excerpt and are not requested;
• 34%: such fields ARE printed in the excerpt (a salary or fee, a governing-law clause, a signature page with names and titles) but are not requested, and the answer leaves them out;
• 15%: a requested field is not stated in the excerpt (a term, a governing-law clause) and takes the stated null convention.

Governing law is requested about as often when it is printed as when it is not (47% stated). Stated values come either from a second excerpt or from the clause inside a single long excerpt. Null requests appear both with no second excerpt and with one that holds a different clause. Neither the presence nor the label of a second excerpt predicts null vs stated (baseline "stated iff a second excerpt is shown": 48.7% vs a 53.2% prior).

How values are taken:

• Parties come only from a strict parse of a two-party opening paragraph ("by and between X (...) and Y (...)") in which both parties carry exactly one quoted label. Three or more parties, nested labels and syndicate language are skipped. So are documents in which another instrument (a grant notice, a cover agreement, another "by and between" paragraph, a vesting schedule) comes before the agreement.
• Labels are given as quoted, without a leading "the" ("the Company" -> Company); the instructions say so. Documents whose label holds anything but letters, spaces, periods and hyphens (stray quotes, OCR residue, digits), or starts with a capitalised "The" that may be part of a name, were dropped (30 documents).
• The agreement date is the single as of / dated / made date in that paragraph. The effective date is asked only when the paragraph says "effective" and no other effective, commencement or start date appears in the shown text.
• The initial term is the verbatim span of the agreement's own term sentence ("the term of this Agreement shall be three (3) years ..."). Sentences about renewal, severance, notice, vesting, non-competition or periods "after" an event are never used. When the term sentence gives only an end ("until December 31, 2009", "until terminated") and no length is printed, the term is null (49 rows).
• Governing law comes only from a governing-law sentence ("governed by / construed in accordance with the laws of"). It is never taken from a state of incorporation. An excerpt used as the governing-law clause always contains that language.
• A second excerpt comes from the same agreement: never from past an exhibit, schedule or annex heading, or another opening paragraph. It must mention a party, a party label or a defined term of the first excerpt, and it must not introduce a party role (Pledgor, Lender, Guarantor ...) or a "this ... Agreement" title that the first excerpt does not have. A signature page must name a party of the first excerpt. It is always a complete paragraph.
• Rows whose copied value (name, label, title) holds OCR residue, such as a run of dashes or underscores, are dropped, for every source.
• Each source agreement belongs to exactly one group, and the same excerpt is asked with different field sets in different rows of that group.
• Excerpts that print an e-mail, phone number or street address were dropped.

Re-admitted rows (2,702 rows). Federal Register and docket rows from the v1 build that an earlier screen had dropped only for sharing a 13-word span with held-out evaluation rows of an older family. A refined screen drops a row if it shares any span with an evaluation item, or at least 8 non-boilerplate 13-word spans with one held-out file; 32 rows stay dropped. The re-admitted rows passed the full decontamination chain again under new ids.

Checks.

• Every answer parses back to its gold object.
• On every contract row the set of answer keys equals the set of requested keys (0 exceptions).
• A question-only classifier predicts null vs stated no better than the per-field prior (lift <= 0 on every field).
• Code checks on every encoded contract row, all 0 exceptions: label characters and no leading "the"; the governing-law excerpt holds governing-law language; a neutral second-excerpt label; no other instrument inside a second excerpt; no effective date beside a defined commencement or start date; every stated term printed in the text.
• The builder read 137 v2 rows against their excerpts; a final 40-row check by the reviewer found 2 errors (OCR dashes inside a name; a second excerpt from an attached pledge agreement), now prevented by the two gates above.
• Blind review round 1 (60 rows) found 6 errors: a term taken from a severance or renewal sentence, an effective date confused with the agreement date, a second instrument before the agreement, a governing-law request that implied null, and conflicting format instructions. The build was regenerated with the rules above.
• Blind review round 2 (60 fresh rows, 57 contract rows) found 1 error (junk in a quoted label) and 4 minor issues (a second excerpt from an attached agreement, a second excerpt mislabelled as the governing-law clause, an inconsistent leading "the", an effective date beside a later commencement date). These fixes, verified by code checks, are in this release.