alea-clause-type

Clause-type tasks over real material contracts filed with the U.S. SEC (EDGAR EX-10 exhibits: employment agreements, equity plans and awards, credit and security agreements, purchase agreements, services and license agreements, leases and others). Every task shows contract text with the section captions removed and asks what kind of provision it is. The gold label comes from the caption the drafters themselves gave the section (the method used to build LEDGAR), mapped to a fixed set of 79 clause types.

kind	rows	task	answer
find_all	8,910	4-9 numbered sections of one agreement; "which sections are X provisions?"	every matching section number (exact set), or none (45%)
presence	8,215	the same kind of excerpt; "does it contain a section dedicated to X?"	yes + the section number, or no (43%)
label_mc	5,600	one provision + 4-6 lettered clause types	the letter, letter + type, or JSON
label_list	5,555	one provision + a list of 6-16 clause types (or the full list of 79)	the type name, or JSON
minimal_pair	1,452	two provisions from different agreements of a near-type pair (assignment vs change of control, termination vs term, indemnification vs limitation of liability, ...)	A or B

Rows are rendered as plain text, Q/A, or chat (about 40 / 30 / 30). About 18% of answers are JSON, with the key named in the instruction (clause_type, type, label, sections, present, ...). About 4% of rows carry one or two short demonstrations whose labels differ from the target. Instruction wordings vary (8-10 per kind).

Sources

Public-record SEC filings from the KL3M data collection: EX-10 exhibits of 10-K and 10-Q filings, 1994-2026. 9,111 agreements from 5,774 filers are used; no filer has more than 4 agreements or more than 0.1% of rows. Agreements that had not been used in earlier training data were selected first. Amendment instruments were left out, because their sections amend other documents. Any section with an e-mail address, phone number, street address, ZIP code or post-office box was never shown. Shown text is a contiguous cut of the filing, never edited; only the caption (and an underline below it) is cut away.

How gold was made

No language model produced any label or answer in this release.

• Sections. Captions were detected by code (numbered captions such as "12. Governing Law.", "Section 5.3 Notices", "ARTICLE VII" + title line, and stand-alone caption lines). A section runs from its caption to the next caption.
• Labels. A hand-written alias table maps exact normalized captions to 79 clause types (for example "Choice of Law", "Applicable Law" and "Governing Law" -> Governing Law). Compound captions ("Governing Law; Jurisdiction") give a label set; those sections appear only in the excerpt tasks.
• Checks (a section failing any of them is never shown or labelled):
  • its body must contain content typical of its type (for example a Notices section must say how notices are delivered);
  • it must not be a single definition;
  • a TF-IDF classifier trained out-of-fold on other agreements must give the caption's type as its top prediction (for compound captions, every type in its top three);
  • leftover captions, sub-captions, signature blocks, page residue, table residue and cut-off list items are detected and the section is dropped;
  • the caption or type name must not appear in caption form inside the body.
• Options and distractors are always other real clause types from the taxonomy, never written text. They exclude:
  • the section's own types;
  • any type whose strict trigger phrase occurs in the passage (for example "change of control" in a termination section);
  • a confusable type whose content check the passage meets. Distractors are sampled so that every type is shown as a wrong option about as often as it is the answer.
• Excerpt tasks.
  • Every shown section is labelled.
  • A type X is asked about only if:
    • no section of another type contains X's trigger phrase (a buried provision);
    • no shown section carries a type confusable with X.
  • "None" answers are real absences: types that occur in agreements of the same kind.

Cue control

The answer cannot be read off the options, the position or the wording alone. These baselines were measured on the released rows (5-fold, folds by agreement):

kind	prior	options-only	position-only	keyword-only	question-only
label_mc	0.194	0.228	0.202	0.245	-
label_list	0.063	0.061	-	0.119	-
find_all (majority)	0.547	-	-	0.553	0.528
presence (majority)	0.573	-	-	0.560	0.566
minimal_pair	0.500	-	0.513	0.562	0.477

Keyword-only means picking the type whose name words occur in the passage. Rows that this baseline solves were thinned until it sat near the prior. Labels are balanced, and no type exceeds 5% of rows.

In addition, a separate model answered 45 multiple-choice items from the options alone, with the provision removed. It scored 24% against a 20% prior, and 18% after removing guesses it attributed to general knowledge of which clauses are common.

Decontamination

LEDGAR, CUAD, MAUD and ContractNLI are built from EDGAR exhibits.

• Hard drops. Every section sharing even one 13-word sequence with any ledgar20 evaluation item, any LegalBench CUAD, MAUD or ContractNLI item (train and test), or the LexGLUE / LegalBench contract items in our benchmark reference was dropped before composition: 31,294 of 360,370 sections. A row-level re-screen dropped 19 more rows.
• Other screens. The standard screens against our other evaluation, benchmark and held-out sets dropped 13,000 more rows. This is mostly shared contract boilerplate, dropped to be safe.
• Note: clause-type classification of SEC contract provisions is the same task type as LEDGAR and the CUAD clause tasks in LegalBench. Models trained on this data no longer measure held-out transfer to that task type on those benchmarks.

Quality review

• Four build rounds were reviewed. Each blind review covered 60 fresh released rows read by a separate model that saw no answer key, and each round's findings were fixed in the generator (by dropping rows, never by editing text):
  • round 1: 4 / 60 rows with a wrong or non-unique answer;
  • round 2: 1 / 60;
  • round 3: 1 / 60;
  • this release: 0 / 60, with 8 style notes.
• The builder also read 30 rows per round.

Fields

id, text (the full training document), family, kind, sub, label (type key of the asked or gold type), source, stratum (filer CIK), gold (JSON: the answer and, where relevant, options, list, section labels), answer, presentation, group_id (one agreement, or an agreement pair), worked, demo, n_tokens (alea-v2b-sw32k tokens). A held-out split of about 1.5% of agreements is not released.

License

CC-BY-4.0. The underlying filings are U.S. public records.