FiledFact-100K — span-grounded, value-verified SEC financial facts

100,000 pairs linking XBRL financial facts to the exact characters that display them in SEC filing text — text[text_start:text_end] == displayed_text holds for every record, and every record deep-links to the exact number in the filing.

Each record contains a passage of clean filing text (financial-statement tables, notes, MD&A) and one fact: the XBRL concept, period, value, unit, and decoded segment dimensions — plus text_start/text_end character offsets locating the displayed value inside the passage, an evidence_url that opens the filing scrolled to that number with a visual marker on it, and the official document URL on sec.gov.

Commercial access & live API

This public sample is free for research under CC BY-NC 4.0. For everything bigger:

• Query instead of download — a free API key (500 calls/month, evidence links included) is self-serve at stockalloy.com/developer; docs at docs.stockalloy.com.
• Full corpus (71 million pairs), quarterly updates, custom slices, point-in-time restatement history, evaluation pilots — commercial license: data@stockalloy.com. From dataset row to source: one click on a fact’s link opens the filing at the exact number, highlighted.

from datasets import load_dataset
ds = load_dataset("StockAlloy/filedfact-100k")
r = ds["train"][0]
r["text"][r["text_start"]:r["text_end"]] == r["displayed_text"]  # True, for every record

Example record

{
  "ticker": "CNO",
  "company_name": "CNO Financial Group, Inc.",
  "form_type": "10-K",
  "filed_at": "2022-02-24",
  "filing_section": "Management’s Discussion and Analysis",
  "concept": "us-gaap#DeferredCompensationArrangementWithIndividualRecordedLiability",
  "concept_label": "Deferred Compensation Arrangement With Individual Recorded Liability",
  "period": "instant:2021-12-31",
  "value": "21200000",
  "unit": "monetary:USD",
  "displayed_text": "21.2",
  "text_start": 294,
  "text_end": 298,
  "scale": 6,
  "sign_in_text": false,
  "dimensions": "[{\"axis\":\"TitleOfIndividualAxis\",\"axis_label\":\"Title of Individual Axis\",\"member\":\"ChiefExecutiveOfficerMember\",\"member_label\":\"Chief Executive Officer\"}]",
  "evidence_url": "https://stockalloy.com/api/evidence/b47accff137cd68f27925b7a3ce918b028534b8cd3e3a10a8190ce7645ec4808/open?cik=0001224608&accession=000122460822000012",
  "text": "In accordance with the terms of the employment agreements of two of the Company's former chief executive officers, certain wholly-owned subsidiaries of the Company are the guarantors of the former executives' nonqualified supplemental retirement benefits. The liability for such benefits was $ 21.2 million and $ 22.0 million at December 31, 2021 and 2020, respectively, and is included in the caption \"Other liabilities\" in the consolidated balance sheet."
}

The sentence says "$ 21.2 million"; the record knows characters 294–298 mean $21,200,000 of Deferred Compensation Arrangement With Individual Recorded Liability, as of 2021-12-31, tagged to the Chief Executive Officer — not the $22.0 million a year earlier, sitting six words away. The link is live: open the filing at that exact number.

Why this dataset is different

1. Exact spans. text[text_start:text_end] == displayed_text holds for every record.
2. Verified values. Each label is checked against the filing's own machine-readable data; measured error rates are published below.
3. Fact-level provenance. Every record carries evidence_url — open the filing scrolled to the value with a visual marker on it — and source_url, the official document on sec.gov.
4. Presentation-layer labels. scale (filing shows "15,992", the fact is 15,992,000), sign_in_text (parenthesized negatives), currency, per-share and percent conventions.
5. Decoded segment dimensions. 45% of records are dimensional (business segment, geography, product, scenario) with human-readable axis/member labels.
6. Extension concepts included. 16.9% are company-specific concepts (ext-*) that most structured-financial datasets drop.

Three tasks, one dataset

FiledFact-100K is sampled at the individual fact level — one verified, grounded fact per row. It is not passage-complete: a passage can hold more facts than the one sampled here, so this release is for exploring and evaluating individual facts, not for passage-level "extract everything" training. A passage-complete release is coming soon.

1. Span grounding — input text + the fact's identity (concept, period, dimensions); locate displayed_text at text_start/text_end.
2. Value normalization — input pre_context + text + displayed_text; predict value_num, unit, scale, and the sign (the traps: "in thousands" headers, parenthesized negatives).
3. XBRL semantic grounding — input pre_context + text + displayed_text; predict concept, period, and dimensions. The most differentiated task — in our internal evaluations, off-the-shelf models performed poorly here (measured results ship with the accompanying model).

Schema (one flat row per pair)

Field	Type	Notes / coverage
id	string	stable pair id
cik / ticker / company_name	string	100% / 99.4% / 99.99%
sic_code	string	98.5% (the remainder are entities EDGAR itself assigns no SIC — trusts, certain funds). Sector/industry are derivable from it via any SIC mapping
accession / form_type / filed_at	string/date	filing identity (10-K/Q, 20-F, 40-F, 6-K + amendments)
fiscal_year / fiscal_period	int/string	filing-level DEI focus (facts may be prior-period comparatives)
source_url	string	100% — the official document on sec.gov
evidence_url	string	100% — StockAlloy resolver that opens the filing scrolled to the value with a visual marker
chunk_id / chunk_type / heading	string	passage identity; table_context 53%, notes 44%, mda + other 3%
filing_section	string	curated human section label for 10-K/10-Q (73%; e.g. "Management's Discussion and Analysis"; null for 20-F/40-F, whose item schemes differ)
pre_context	string	visible text immediately before the passage (99.2%) — often carries the table caption and scale cue
text	string	clean filing passage (tables linearized with | cell separators)
concept / concept_label / is_extension	string/bool	XBRL concept + human label (100%)
period	string	canonical: instant:YYYY-MM-DD or duration:YYYY-MM-DD..YYYY-MM-DD — type and dates are parsed from it directly
value / value_num / unit / unit_class / currency	string/double	value is the canonical value as a string and preserves precision. value_num is a convenience numeric field for analysis
displayed_text / text_start / text_end	string/int	the value as rendered + exact char span (verified for 100% of rows)
scale / sign_in_text / scale_cue_in_text	int32/bool	scale = power of ten the display is compressed by (3 = "in thousands": shows 69,542, means 69,542,000; 99.8%); sign_in_text = passage displays the number as negative (parens/minus) — compare with the sign of value_num; cue presence 63%
dimensions / dims_count	JSON string/int	decoded [{axis, axis_label, member, member_label}]
confidence	float	the aligner's certainty for the text↔fact pairing (0.75–0.99). Note: our audit found semantic-label errors are NOT concentrated in lower tiers — treat the ~2% risk as uniform

Splits are company-disjoint (train 94,084 / validation 5,916 by deterministic CIK hash; selection seed 20260706). The 150 companies of our eval benchmark are excluded from both splits.

Schema note (2026-07): scale is now typed int32 (was float64 — a null-coercion artifact). Redundant columns were removed: sector/industry (derive from sic_code), filing_part/filing_item (superseded by the curated filing_section), and period_type/period_start/period_end (all encoded in the canonical period string).

Coverage

5,668 companies · 54,452 filings · fiscal years 2016–2027 · US GAAP and IFRS (20-F) filers · multiple currencies.

Data quality & known limitations

• Every row passed numeric value-verification at selection (the corpus's value_verified filter).

• Measured label error: an independent audit of 1,000 random parent-corpus pairs found ~2% semantic errors — records where the span and arithmetic are correct but the label misidentifies the fact. The types found, with measured rates:

  • Wrong concept (~1%) — a real value tagged as the wrong line item (e.g., a total labeled as one of its components).
  • Wrong period (~0.5%) — the value attached to the neighboring column or year: the twin-value trap in "was $21.2 million and $22.0 million ... respectively" sentences and multi-year tables.
  • Wrong sign (~0.2%) — the normalized value's direction flipped, beyond the documented XBRL outflow conventions.
  • Wrong scale (~0.1%) — normalization off by a power of ten (a "in thousands" table treated as millions, or vice versa).

  Treat ~2% as an upper bound for this sample: its construction excludes the main error sources — single-occurrence spans remove the ambiguous-alignment class, and every parseable row passed an independent displayed×scale↔value reconciliation (99,848/99,848) before release.

• Filer-faithful values (rare filer errors included): every value, scale, and unit reproduces the filer's own iXBRL tags. We swept all 100,000 rows for implausible magnitudes and traced every anomaly to its source ix:nonFraction node in the original filing: 11 rows (~0.01%) carry the filer's own mis-tags — count columns that inherited a dollar column's scale (claims/units/barrels ×10³–10⁶), percentages the filer left un-normalized (5% tagged as 5 rather than 0.05, a pattern that appears across EDGAR), one 100× ratio, and one CAD/USD unit mislabel. The same wrong values appear in SEC's official data for those filings. The full registry ships in this repo (audit/mistags.json).

• Sign conventions are XBRL's, not errors: ~8% of pairs display a parenthesized (negative) number while the value is positive (outflow concepts are positive by definition); some values are negative with no visible sign. sign_in_text + value_num give you both sides.

• Scale cues can live outside the passage for a minority of records; the scale column is derived numerically, not from the cue text.

• Sampling: this sample keeps only pairs whose displayed value appears unambiguously in the passage — which is what makes every span exact, and which under-represents repeated-value cases (basic-EPS-equals-diluted rows, totals matching another line, small round numbers, per-share facts). Dash-for-zero displays were excluded.

• Text cleanliness: under 1% of rows retain small HTML fragments in text (imperfect markup stripping); their spans and values remain exact and verified.

• The fine-tuned model published alongside this dataset was trained on an earlier text extraction, not on this exact sample.

• This public release is a demo/research sample, not a benchmark: its labels are public. A private blind benchmark (hidden labels, hard-case selection, decomposed scoring) is available for evaluation pilots — see below.

Benchmark

In preliminary internal evaluations, off-the-shelf models could usually read the displayed number but performed poorly at identifying which financial fact it is — the concept, period, and segment. Fine-tuning on FiledFact is what teaches that. The fine-tuned model and the full measured benchmark are published: StockAlloy/filedfact-qwen3-8b (Apache 2.0 — fact-F1 0.295 → 0.787 on the held-out benchmark; the model card carries the complete comparison incl. gpt-5.5).

Evaluation protocol (also the fine-tuning format)

Our benchmark evaluates whole-passage extraction: the model receives one filing passage and must return every fact in it. The eval set is a held-out group of 150 companies (excluded from this sample and from training); its labels are not published.

Input — the record's form_type, filed_at, pre_context, heading, and text, rendered with this exact prompt:

You are a financial data extraction system. Below is a passage from an SEC filing ({form_type}, filed {filed_at}). Extract every XBRL-tagged financial fact that appears in the passage.

Return ONLY valid JSON in this exact shape:
{"facts": [{"concept": "...", "value": 0, "unit": "...", "period": "...", "dimensions": [{"axis": "...", "member": "..."}]}]}

Rules:
- concept: the US-GAAP/IFRS XBRL concept name if you know it (e.g. "RevenueFromContractWithCustomerExcludingAssessedTax"); otherwise a precise CamelCase label.
- value: the numeric value in FULL units. Expand the table scale: if the table is "in thousands" and a cell shows 1,634 the value is 1634000. Use negative numbers for values shown in parentheses.
- unit: ISO currency code (USD, EUR, ...) for monetary values; "shares" for share counts; "pure" for ratios.
- period: "instant:YYYY-MM-DD" for point-in-time balances, or "duration:YYYY-MM-DD..YYYY-MM-DD" for flows over a period.
- dimensions: the slice the value belongs to (business segment, geography, product, plan type, share class, ...) as axis/member pairs. Use [] for company-wide totals.

CONTEXT IMMEDIATELY BEFORE THE PASSAGE (may contain the table caption and scale, e.g. "in thousands"):
{pre_context}

SECTION HEADING: {heading}

PASSAGE:
{text}

Unit mapping (dataset ↔ prompt): dataset rows carry class-prefixed units (monetary:USD, per_share:USD, shares, pure, other), while the extraction prompt above asks for the short form. The mapping is: strip the class prefix — monetary:USD and per_share:USD → USD, shares → shares, pure → pure. The unit_class column preserves the class if you need to distinguish per-share monetary values.

Worked example — for the CNO record shown above, the prompt's variable part renders as:

CONTEXT IMMEDIATELY BEFORE THE PASSAGE (may contain the table caption and scale, e.g. "in thousands"):
7.6 million and such receivables were $ 14.8 million. These estimates are subject to change when the associations determine more precisely the losses that have occurred and how such losses will be allocated among the insurance companies. We recognize

SECTION HEADING: Guarantees

PASSAGE:
In accordance with the terms of the employment agreements of two of the Company's former chief executive officers, certain wholly-owned subsidiaries of the Company are the guarantors of the former executives' nonqualified supplemental retirement benefits. The liability for such benefits was $ 21.2 million and $ 22.0 million at December 31, 2021 and 2020, respectively, and is included in the caption "Other liabilities" in the consolidated balance sheet.

and the expected output is both facts in the passage — the current value and its prior-period twin, each with the right period and the individual-axis dimension:

{"facts": [
  {"concept": "DeferredCompensationArrangementWithIndividualRecordedLiability", "value": 21200000, "unit": "USD", "period": "instant:2021-12-31", "dimensions": [{"axis": "TitleOfIndividualAxis", "member": "ChiefExecutiveOfficerMember"}]},
  {"concept": "DeferredCompensationArrangementWithIndividualRecordedLiability", "value": 22000000, "unit": "USD", "period": "instant:2020-12-31", "dimensions": [{"axis": "TitleOfIndividualAxis", "member": "ChiefExecutiveOfficerMember"}]}
]}

Note what the model had to get right: expand "21.2 million" to 21200000, attach each value to its own period from a shared "respectively" sentence, and carry the CEO dimension on both.

Output — the facts JSON above. Scoring — predicted facts are matched to gold facts and each field is scored separately: fact recall/precision (value+period match), concept accuracy, unit accuracy, dimensions F1, and full-fact rate (everything correct at once). No partial credit for a correct number attached to the wrong concept, period, or segment — that failure mode is the point of the benchmark.

Rows in this dataset are one fact each; to reproduce the passage-level training format, group by chunk_id (and see the sampling note below — this public sample does not contain every fact of every passage).

Intended uses & limitations

Uses:

• Fine-tuning extraction models (passage → structured facts)
• Grounding & citation research (teach models to point at their evidence)
• Table understanding (row/column/scale association on real financial tables)
• RAG evaluation with verifiable provenance
• Financial-NLP benchmarks and error analysis (per-field quality columns)

Limitations: verified+grounded subset (biased toward cleanly rendered facts); passage text refers to the filer generically ("the Company") — use the identity columns for company context; filing-level fiscal focus vs per-fact comparative periods; US SEC filers only; not investment advice.

The full corpus

This sample is a slice of FiledFact: 71 million span-grounded pairs, 6,100+ companies, 94,700 filings (2016–2026), quarterly updates, plus a point-in-time as-reported restatement history. For document-AI, OCR, and VLM teams we also run private blind evaluation pilots — hidden labels, hard-case selection, decomposed scorecards (value / row / column / period / scale / sign / segment / concept / evidence).

Commercial licensing, custom slices, and evaluation pilots: data@stockalloy.com · stockalloy.com/datasets · or open a discussion here.

License & citation

Sample: CC BY-NC 4.0, applying to StockAlloy's compilation, annotations, alignments, and derived metadata. The original SEC filings are publicly accessible through EDGAR and are not owned by StockAlloy; nothing here restricts independent access to them. Commercial use of the annotations and alignments requires a license — see above.

Evidence links are intended for human verification and light programmatic use — please don't bulk-crawl them. For bulk evidence access or API use, contact us for a licensed endpoint.

@dataset{stockalloy_filedfact_2026,
  title   = {FiledFact-100K: Span-Grounded, Value-Verified SEC Financial Facts},
  author  = {StockAlloy Research},
  year    = {2026},
  url     = {https://huggingface.co/datasets/stockalloy/filedfact-100k}
}