Peter-A-P/smallprint-sec-extraction

Datasheet: smallprint SEC filings extraction corpus

Structured extraction from the primary financial statements of SEC 10-K and 10-Q filings into fifteen fields, each labelled with the XBRL fact the company itself filed. Built by smallprint on 2026-09-22. Every number below was written by the build; none was edited by hand.

Composition

7,874 items from 1,600 selected companies, filings dated 2022Q1 to 2026Q2, forms 10-K, 10-Q. Of 19,574 filings tried, 10,497 became items before the split and 9,077 were dropped with a named reason (below). The split then drops a further 2,623 training-pool filings dated after the cutoff.

Each item is one filing: text, the model's input (the cover page, the income statement, the balance sheet and, in an annual report, the audit signature, located and rendered from the filing's HTML with hidden inline XBRL removed); truth, the fifteen labels; and context, the reporting scale and the other periods' values of each line, which the grader uses to name a wrong-period reading. The median input is 6,561 characters.

Splits

Companies, not filings, are assigned to pools, so no company appears in both training and test. The cutoff, 2025-01-31, is the latest documented training cutoff of the base models; test-pool filings after it are the headline set, and before it the contamination comparison. Test filers appear in both test sets by design.

Split	Filings	Companies	10-K	10-Q
train	5,060	585	1,145	3,915
validation	713	88	161	552
test_pre_cutoff	1,396	165	287	1,109
test_post_cutoff	705	144	186	519

Fiscal periods: FY 1,779, Q1 2,079, Q2 2,009, Q3 2,007.

Where the labels come from

Numeric labels are XBRL facts from the SEC companyfacts API for the filing's own accession, for the period the report covers; cover-page labels are the filing's own inline XBRL tags. No model produced or checked any label. A filing is kept only when every label is printed in the section of the input its field is read from, within the grader's tolerance after the printed scale, so the task is extraction and not inference. The one exception is diluted weighted shares: when the income statement prints no share-count line at all, that label is null and the filing is kept (768 filings). The rules and their evidence are in docs/data.md.

Label coverage

Field	Labelled	Items
period_end	100%	7874
fiscal_period	100%	7874
revenue	100%	7874
cost_of_revenue	62%	4905
operating_income	88%	6931
net_income	100%	7874
eps_basic	88%	6957
eps_diluted	85%	6711
shares_diluted	79%	6245
total_assets	100%	7874
total_liabilities	89%	6987
cash_and_equivalents	100%	7874
stockholders_equity	100%	7874
auditor_name	22%	1770
state_of_incorporation	96%	7548

What is missing, and why

Reason	Filings	Share of those tried
truth_missing:revenue	3,214	16%
bank	1,140	6%
no_income_statement	949	5%
not_fetched	693	4%
no_balance_sheet	470	2%
unlocatable:auditor_name	403	2%
no_period	389	2%
unlocatable:revenue	344	2%
unlocatable:shares_diluted	314	2%
unlocatable:operating_income	191	1%
unlocatable:total_liabilities	184	1%
unlocatable:stockholders_equity	126	1%
truth_missing:stockholders_equity	103	1%
unlocatable:cash_and_equivalents	92	0%
unlocatable:eps_basic	90	0%
unlocatable:cost_of_revenue	88	0%
unlocatable:state_of_incorporation	63	0%
truth_missing:net_income	58	0%
unlocatable:total_assets	46	0%
truth_missing:cash_and_equivalents	35	0%
unlocatable:period_end	29	0%
unlocatable:eps_diluted	23	0%
truth_missing:total_assets	14	0%
unlocatable:net_income	8	0%
mixed_scale	7	0%
truth_invalid:cash_and_equivalents	4	0%

Underrepresented on purpose. Banks and savings institutions (industry codes 6021, 6022, 6029, 6035, 6036) are excluded: they report no single revenue line to label. Funds, trusts, business development companies and SPACs mostly drop out because they print no income statement of the kind the task reads, or no revenue. Filers whose statements print no reporting scale are dropped rather than labelled by guessing it. The corpus therefore underweights financial companies and very small or pre-revenue ones, and a score on it says nothing about extraction from their statements.

Collection

Fetched from SEC EDGAR under its fair-access policy: a declared contact address in every request, paced below ten requests a second, every response cached with its date and digest so the corpus rebuilds offline without refetching. Company selection is a keyed hash, seed 20270405. Rebuild with:

smallprint data build --first 2022Q1 --last 2026Q2 \
  --cutoff 2025-01-31 --companies 1600 --seed 20270405

Licence

SEC filings and XBRL data are public records of the US government. The licence under which this corpus is published is decided before publication and recorded here.

Checksums

13dc546b3d616ac9098c21a37ab784c9a285fb96bd5aa8e84ec0e461a13a0ed3  items.jsonl
82fb8d2506a9467bbe5cc22aa02bb71eb1baa597a47187a23c42405bc2745228  dropped.jsonl
242ce0ddb65e8ae01bf393623986004582fbe8260a30f0a6760743e2b2434cd6  report.json