alea-institute/alea-ex21-diff

EX-21 Subsidiary List Diff

This dataset contains multi-step reasoning tasks over real, public-record subsidiary lists (Exhibit 21) that companies filed with their annual reports (Form 10-K) at the U.S. SEC. Each document shows two or three subsidiary lists of the same company from different filing years. Each list is reproduced verbatim from the filing, labeled with its filing date, and the lists appear in random order. The questions ask the reader to compare the lists, count, apply set operations and track change over time.

No language model wrote any part of this data. Every question, answer and worked explanation is produced by code from the parsed lists, and every answer is checked against the lists shown.

Source

• Origin. Exhibit 21 subsidiary lists from the KL3M EDGAR 10-K corpus (public-record SEC filings), filed 2001-2026.
• Which lists are used. Only lists that are complete markdown tables with:
  • one header line;
  • 6-40 entity rows, each with a recognised jurisdiction;
  • no blank jurisdiction cells, no inner headings, no wrapped or duplicated names;
  • no footnote marks, percentages or status notes ("formerly", "dissolved", ...) inside names.
• Pairing. One document group per company. A pair is two lists from different filing years (mostly 1-5 years apart) that share at least 40% of their names.
• Pairs that are skipped. A pair is skipped if a name in one list is only near-equal to a name in the other, because that would make "added / removed" ambiguous. Near-equal means: edit distance <= 3, a prefix, the same core name, the same words reordered, or one word more or less.
• Name matching. Names count as the same only under a simple canonical form: case, punctuation and whitespace are ignored, Inc. = Incorporated, Ltd. = Limited, Corp. = Corporation, Co. = Company, LLC = L.L.C. When two lists print the same entity differently, the instructions say so.
• Caps. No industry (4-digit SIC) is above 5% of documents. No company is above 0.1%.

Tasks

Each document has 1-3 numbered questions.

• added / removed. Which subsidiaries appear in the later list but not the earlier one, or the reverse. Names are given as printed, in list order; the answer is "None" when there are none.
• count_change. The number of entities in each list and the signed change, stated in words ("decreased by 3").
• jur_count_change. The same as count_change, for one jurisdiction.
• moved. Whether an entity's listed jurisdiction changed, and from what to what. These come with no-change twins.
• membership. Whether an entity is listed under exactly a given name in both filings, only the earlier, only the later, or neither.
  • "Neither" names are near-name variants verified absent from both filings: a suffix swap (at most half), a numeral change, or one word swapped for a word from another listed name.
  • When a listed entity differs from the asked name only in its entity-type suffix, the answer names that listed entity.
• abstain. Tempting questions the lists cannot answer, about a real listed entity: acquisition date, ownership percentage, why it was dropped, headcount, formation date, officers, revenue. The answer is "The lists do not say." Absence of the information is checked by code over both passages.
• Three-filing state tracking. Entities dropped then listed again; entities in all three lists; the first filing to list an entity; entities in the first list that are missing from both later lists.

Every question needs 2-4 dependent steps, for example: order the filings by date, look up or count in each list, then compare or subtract. The steps are recorded in gold.steps.

About 40% of documents carry short worked steps before a final Answer: line. Each worked document has a direct-answer twin of the same question in another document of the same group. Twins and contrastive pairs are always split across documents of a group:

• added vs removed;
• moved vs unchanged;
• two jurisdictions;
• an abstention question vs an answerable question about the same entity.

About 4% of documents include one worked demonstration from another company. Presentations vary: plain instruction (40%), Q/A markers (30%) and chat turns (30%).

Size and composition

• 7,231 training documents, about 5.2M tokens (alea-v2b-sw32k). Median 634 tokens per document, max 2,370.
• 1,602 companies.
• Group types: two changed lists 80%; two unchanged lists 9.5% (so "none" is a valid answer); three lists 10.6%.
• Question counts: membership 4,667; jurisdiction count change 2,014; added 1,805; removed 1,788; count change 1,366; abstain 1,169 (8.2% of questions); first appearance 586; in-all-three 336; dropped for good 237; moved 113; relisted 27.
• A held-out probe split of 78 documents (about 1.5% of companies) is kept separate.

Quality checks

• Independent re-check. A separate script re-reads the lists from the rendered text and recomputes every set, count, membership and state answer. It found 0 mismatches over 10,857 questions.
• Shortcut baselines. Question-only, name-only and position-only classifiers (5-fold, grouped by company) stay within about 3 points of the majority-class prior for every labelled question type, except a noisy question-only baseline on the small moved type (113 questions).
• Label balance.
  • Count changes: increase 37%, decrease 40%, no change 23%.
  • Membership classes: 20-29% each.
  • The position of the true option in membership questions is uniform.
• Decontamination. Documents were screened against benchmark and held-out evaluation items, and documents overlapping them were removed.
• Manual review. The builder read the final documents against their sources. A blind review of 60 documents found 1 ambiguous question wording, which has been fixed, and soft issues with membership near-names, which have been addressed. A second independent blind review of 60 new documents re-derived every answer from the lists: 0 wrong answers, 2 style notes (a translated legal form, "Corporation" vs "Kabushiki Kaisha", counted as different names; one awkward question wording). Twelve documents whose presentation label doubled the answer label ("Answer: Answer:") were removed afterwards.

Limitations

• Passages are verbatim, including source typos, letter-spaced headings and extra columns. Answers follow the printed names exactly.
• "Same entity" is defined by the canonical name form above, not by legal identity. A renamed subsidiary appears as one removal and one addition (pairs where this would be ambiguous are excluded).
• EX-21 lists may omit non-significant subsidiaries. Questions are about what the lists show, not about the company's full corporate structure.