bowang0911/sec-10k-item-hn

SEC 10-K Item-Section Hard Negatives

Query / positive-passage / hard-negative triplets built from the narrative sections of SEC Form 10-K annual reports. Positives and negatives are real filing prose; only the query is model-generated.

from datasets import load_dataset
ds = load_dataset("bowang0911/sec-10k-item-hn")   # 7,548 rows, single train split

7,548 rows from 3,911 filings / 3,892 distinct companies. Source: one 10-K per CIK (most recent), drawn from the SEC Financial Statement Data Sets index and fetched from EDGAR; up to 3 rows per filing. 7,239 filings were processed; 41% yielded no eligible section (typically small filers whose headings are not detectable as such).

rows	7,548 (all queries unique)
Items	1A 3,949 · 1 1,841 · 7 1,635 · 7A 69 · 2 20 · 1C 17 · 3 11 · other 6
negatives per row	6.9 mean, 4 minimum
negatives from the positive's own Item	61%
lexical margin	mean −0.101, ceiling +0.10
query length	p50 16 words, p90 21, max 26
fiscal years	2025 6,023 · 2024 514 · 2022 510 · 2023 408 · 2026 89 · other 4

12,365 rows were built and 2 failed judging. Of the remaining 12,363: 1,167 were dropped as corrupt text, 68 because another row already used the same positive, and 3,580 as judgement calls about difficulty or query quality. The two large sets are itemised below, since what was thrown away is the main thing you need in order to trust what is left.

Construction

Each row starts from a level-1 heading inside a Regulation S-K Item (Item 1 Business, 1A Risk Factors, 7 MD&A, ...). The heading is rewritten into a question; the section it labels becomes the positive; sibling sections from the same filing become negatives, preferentially from the same Item so the distractor shares the disclosure context.

stage	kind	purpose
Item spans	rule	last-occurrence dedup + longest-increasing-subsequence over the Item taxonomy, to separate body from table of contents
running headers	rule	drop lines repeated ≥3× (filers print bare Item 1A on every page, which mis-assigns spans)
boilerplate	rule	drop prescribed sections: non-GAAP reconciliations, forward-looking statements, internal-control reports
tables	rule	drop passages whose token stream is >15% numeric (stripped tables read $ 34,550 % $ 31,370 %)
rephrase	LLM	heading → question, 1 of 4 style forms, rejected if it reuses >60% of the passage's vocabulary or exceeds 26 words
lexical margin	rule	require overlap(q, positive) − max overlap(q, negative) ≤ +0.10, so BM25 alone cannot rank the positive first
co-relevance, pass 1	LLM	drop any negative that also answers the query
co-relevance, pass 2	LLM	drop any negative that is on the query's specific subject, whether or not it answers in full; also flag the query as vague
co-relevance, pass 3	LLM	judge every remaining negative on its own, then confirm each flag by majority vote of three models
duplicate negatives	rule	drop any negative sharing a long verbatim run with the positive
risk summaries	rule	drop negatives that enumerate the filer's whole risk-factor list

Item boundaries are disclosure-obligation boundaries, not topic boundaries, so a sibling section frequently covers the same subject as the positive. Both judging passes act on the negative, not the row: a row whose fourth negative is co-relevant keeps its query, its positive, and its other negatives.

pass	criterion	dropped
1	negative also answers the query	8,058 of 97,931 (8.2%)
2	negative is on the same specific subject	8,489 of 89,873 (9.4%)
3	one call per passage, majority-of-three confirmation	3,091 of 81,370 (3.8%)
all		19,638 of 97,931 (20.1%)

Pass 3 exists because passes 1 and 2 each sent the positive and all negatives in one call and asked for a list of relevant indices. That framing biases the model toward returning a short list. Judging one passage per call, with the same model and the same criterion, flags roughly twice as many negatives — 21.6% against 9.2% on a 218-negative sample — and on a hand-checked row the batched call returned nothing at temperature 0 across three runs while the per-passage call found the co-relevant negative.

Per-passage judging alone over-fires, so each flag is then re-judged by three models that must quote the span making the passage relevant, and the majority decides. The confirmers have opposite biases of similar size, which is why the vote beats any of them alone: Sonnet 4.5 wrongly confirms passages sharing a broad category with the question (a labour-disruption risk factor for a question about legal and regulatory compliance risks), while gpt_5_2 and Opus 4.5 wrongly reject passages that cover the subject from an unexpected angle (segment revenue percentages for a question about segment performance). Opus is only called on the 30% of flags where the first two disagree.

Pass 3 runs on what pass 2 left, so the three drop sets compose. This matters: passes 2 and 3 agree on only about a third of their drops (Jaccard 0.34 over 300 negatives), and hand-checking the disagreements found real co-relevant negatives on both sides. Pass 3 catches passages the batched call skipped; the batched call catches long passages where a single sentence near the end carries the figures. That position effect is the core difficulty — passages run to several hundred words and the co-relevant sentence is often the last one, so both pass-3 prompts instruct the model to read to the end.

Pass 2 exists because pass 1's "does it answer" bar is too high. A passage that discusses the same specific risk, obligation, programme or transaction is co-relevant even when another passage covers it better, and pass 1 kept those. Pass 2's prompt also refuses the inverse error explicitly — shared company, fiscal year, Item or broad theme is not enough, so a question about currency controls limiting dividends does not make a passage on national economic growth relevant.

Columns

• query, document, hard_negatives — the triplet. document and each negative are prefixed {Company} — FY{year} Form 10-K.
• meta_item, meta_heading — which Item the positive came from, and its heading.
• meta_lex_margin — the margin above, recomputed over the negatives that actually ship. Sort descending to find the most keyword-solvable rows.
• meta_same_item_negs — negatives drawn from the positive's own Item. Floor is 2; rows at the floor have the weakest distractors.
• meta_dropped_corelevant — negatives removed across both judging passes.
• meta_rejudge_dropped — negatives removed by pass 2 alone.
• meta_query_spec_idf — corpus IDF of the rarest content token the query shares with its positive. Low values mark queries built from generic risk vocabulary. Diagnostic only: it does not predict false negatives (see below).
• meta_title_overlap — heading-vs-own-body vocabulary overlap (not heading-vs-Item-title).
• meta_query_overlap — query-vs-positive vocabulary overlap after removing the company name and fiscal year, which the query is required to state.
• meta_flags — the judgement-flag column. Empty for every published row, since flagged rows are excluded; retained so the schema matches the export script.

Negatives deleted from surviving rows

Two classes are labelling errors rather than hard cases, so they are removed from every row that carries them; the row survives on its remaining negatives.

class	rows	negatives	why
duplicate of the positive	108	112	the negative restates the positive, so the loss pushes the query away from text it also pulls it toward. One case was byte-identical; another was the verbatim first third of a 437-word positive.
risk-factor summary	147	147	an Item 1A summary enumerating every risk the filer discloses (The following is a summary of the principal risks that could…) is co-relevant to nearly any Item 1A query. No positive in the corpus is such a passage, so nothing is lost.

The duplicate test is longest contiguous shared word-run ≥ 50% of the shorter passage, after whitespace normalisation. Set-overlap measures were tried and rejected: two passages describing different private placements in the same house boilerplate reach 0.64 by 8-gram containment and 0.91 by token containment while being a perfectly good query/negative pair. Verbatim reuse — which is what section trimming produces when a filer repeats text across Items — instead shows up as a single long run. After filtering, the corpus-wide maximum is 0.48.

Rows removed as corrupt (1,167)

defect	rows	why
numeric_positive	932	positive is >10% numeric tokens — a stripped table, e.g. OPERATING INCOME 1 (18.8) Adjusted EBITA 1,517.1 1,566.6
bad_entity	90	registrant name captured a parenthetical or sponsor line: (An Electric Membership Corporation)
banner_heading	24	heading restates the Item banner (Item 7. Management's Discussion…)
whitespace_loss	22	HTML extraction lost inter-element spaces: Wemaybesubjecttounknownorcontingentliabilities
positive_not_relevant	68	the judge would not call the positive relevant to its own query
caption_heading	19	heading is a table units caption: (Amounts in Millions, Except Per Share Amounts)
no_negatives	26	every negative was deleted as co-relevant, duplicate or risk-summary
no_fiscal_year	7	dei:DocumentFiscalYearFocus did not parse
unanswerable_quantity	6	query asks "how many / how much" and the positive contains no digit — the figure was lost with the table

Rows can carry more than one defect, so the column sums above the row total.

Rows removed as judgement calls (3,580)

These rows are not corrupt. They were withheld because something about their difficulty or their query makes them a worse bet than the rows that shipped.

reason	rows	what it means
lex_solvable	2,240	after judging deleted co-relevant negatives, lex_margin > +0.10 — the positive is the token-overlap argmax outright, so BM25 may solve the row. The build applied this cap before the deletions, so it never bound on the shipped set; re-applying it post-hoc is what surfaces these.
rejudge_below_min	1,376	judging left the row under 4 negatives or under 2 same-Item negatives. The remaining negatives are sound; the row is just thinner than the floor.
vague_query	1,021	pass 2 judged the query general enough that several passages answer it about equally well. These are the rows most likely to contain a false negative.
group_banner_heading	198	the heading is a risk-section category divider (Risks Related to Our Financial Position and Need for Additional Capital), not a risk factor. The question inherits the whole category, so other passages in it answer as well as the positive does. Detected as a Risks related to … prefix with no verb — real 1A headings are full sentences.
below_min_negatives	9	under the floor for a reason other than judging.

Rows can carry more than one reason. Note that these are not one failure mode: lex_solvable rows are too easy, vague_query and group_banner_heading rows are false-negative prone, and rejudge_below_min rows are merely thin — a row with 3 sound negatives is usable if you do not need a fixed count. Withholding all of them is the conservative choice, not the only defensible one.

Filtering philosophy

False negatives in a contrastive denominator are not neutral — they train the model to push apart a pair that should be close, which is why undenoised top-k hard negatives can underperform random negatives. Hence the two judging passes, and hence their granularity: they act on the negative, so a co-relevant fourth negative costs one distractor rather than a whole row.

Row-level deletion is a sharper instrument and is used sparingly, because the hardest legitimate negative is a passage on the same topic that does not answer the specific question — exactly what an eager false-negative filter removes first. Filter hard enough and what survives is easy topic-shift negatives, and the model learns topic matching instead of the precision the dataset exists to teach. The lex_solvable count above is the visible symptom of the opposite risk: filtering for co-relevance makes rows easier, not harder.

Note also that most rows reading as false negatives are really underspecified queries — which is why pass 2 reports vague_query separately. For a query about how patent disputes affect a company's finances, a negative about the cost of defending patents is not clearly irrelevant; the question is too vague to have one answer. Sharpening the query removes the false negative and keeps the hard negative; dropping the row loses both. That is a fix to the rephrase prompt, and it has not been made.

Known issues

• A hand audit of 10 random rows found 1 probable false negative and 1 borderline case. The same 10-row draw before pass 2 gave 5 false negatives and 2 borderline, so pass 2 moved the rate materially. The estimate is still 10 rows on one seed, by one reader, and it is not a substitute for a labelled eval. The surviving probable case was the risk-summary passage that motivated the filter above; the borderline one is a marketing-strategy passage against a market-share query.
• Lexical gates do not find these rows. An IDF-specificity floor was calibrated against hand-labelled rows and did not separate them (a floor of 3.0 keeps 3 of 5 known-bad rows while dropping a clean one); a rule dropping negatives that share the query's rarest topic token deletes sound negatives wholesale, because for a single-country filer that token is the country. Both are implemented in the export script and both are off by default. The failures are semantic — terrorism/terrorist and funding/financing are not linked by any stemmer — which is why the fix was a second judging pass, not a filter.
• Item 1A is 52% of rows. The candidate order round-robins across Items, but many small filers only produce detectable level-1 headings inside Risk Factors.
• Embedded table fragments survive in ~4% of positives. The numeric filter is a whole-body average, so mostly-prose passages with a table fragment inside pass. A per-line filter in the extractor is the real fix and has not been done.
• Cross-row duplicate detection catches only exact matches. Rows whose positives are byte-identical after whitespace normalisation are deduped, but the build's near-duplicate guard runs on untrimmed bodies, so a filer repeating the same prose across Items at different lengths can still produce two rows with near-identical positives.
• The judging tie-break deletes rather than preserves. Every pass is told to mark a passage relevant when genuinely unsure, and a passage marked relevant is removed from the negatives. The bias is therefore toward deleting borderline hard negatives, not toward leaving false negatives in. Judge accuracy on SEC prose is measured only by hand audit. Before pass 3, 2 of 20 rows drawn under two seeds carried a probable false negative; after pass 3, a fresh draw of 10 rows carried none, with one vague query left over.
• Cross-row leakage is measured; only the harmful part is removed. Before dedup, 1,438 of 8,102 rows (17.8%) had a positive that also appeared as a negative in some other row — 1,560 slots, 1,516 of them inside the same filing. Those are mostly benign: the other row has a different heading, and that row's own judging pass already weighed the passage against its own query. What is not benign is two rows whose positives are the same prose, which a filer repeating text across Items can produce; with in-batch negatives each row's positive then trains against the other's. The export now keeps one row per distinct positive text, dropping 71 rows. No BM25 leak check has been run.