FiledFact-Passages - passage-complete, span-grounded SEC fact extraction

5,698 SEC filing passages containing 101,899 financial facts - every fact grounded to its exact characters, every other numeric span labeled as a non-target, and every fact linked to the exact number in the original filing.

Each row contains one clean filing passage (financial-statement tables, notes, or MD&A) and every tagged fact displayed in it: concept, normalized full-unit value, period, unit, decoded dimensions, exact character offsets, and an evidence_url that opens the filing at that number. Years, dates, references, prose percentages, and untagged amounts are labeled in non_target_spans.

• Sibling dataset: FiledFact-100K provides one grounded fact per row.
• Companion model: FiledFact-Qwen3-8B.

from datasets import load_dataset

ds = load_dataset("StockAlloy/filedfact-passages")
row = ds["validation"][0]

for fact in row["facts"]:
    assert row["text"][fact["text_start"]:fact["text_end"]] == fact["displayed_text"]

	Count
Passages	5,698
Companies	4,878
Tagged facts	101,899
Non-target numeric spans	50,658

The task

Input: one filing passage, with optional nearby context.

Output: every tagged financial fact displayed in the passage.

A validation example is a goodwill-by-segment table whose pre_context says “In thousands”:

Engineering
|   |   | Specialty
Health Care
|   |   | Information
Technology
|   |   | Total
$ | 11,918 |   |   | $ | 2,398 |   |   | $ | 7,831 |   |   | $ | 22,147 |

Its four facts:

{"concept": "us-gaap#Goodwill", "value": "11918000", "unit": "monetary:USD",
 "period": "instant:2023-04-01", "text_start": 97, "text_end": 103,
 "dimensions": [{"axis": "StatementBusinessSegmentsAxis",
                 "member": "ext-0000700841#EngineeringMember"}]}

{"concept": "us-gaap#Goodwill", "value": "2398000", "unit": "monetary:USD",
 "period": "instant:2023-04-01", "text_start": 118, "text_end": 123,
 "dimensions": [{"axis": "StatementBusinessSegmentsAxis",
                 "member": "ext-0000700841#SpecialtyHealthCareMember"}]}

{"concept": "us-gaap#Goodwill", "value": "7831000", "unit": "monetary:USD",
 "period": "instant:2023-04-01", "text_start": 138, "text_end": 143,
 "dimensions": [{"axis": "StatementBusinessSegmentsAxis",
                 "member": "ext-0000700841#InformationTechnologyMember"}]}

{"concept": "us-gaap#Goodwill", "value": "22147000", "unit": "monetary:USD",
 "period": "instant:2023-04-01", "text_start": 158, "text_end": 164,
 "dimensions": []}

The displayed 11,918 means 11918000 because the scale cue is in the nearby context. The segment cells carry dimensions; the total has an empty dimensions list.

Some passages intentionally contain no target facts (labeled abstention|0 in sample_stratum).

What passage-complete means

Each passage contains:

• facts - the financial facts to extract (the numbers the filer tagged in the filing).
• non_target_spans - every other numeric span, including years, dates, references, prose percentages, and untagged amounts.

A visible numeric token is assigned to one of those two groups. Co-located XBRL facts are kept separately when the filer attached multiple concepts, periods, or dimensions to the same displayed number.

Values and evidence

• displayed_text is exactly what appears in the passage.
• text_start and text_end are half-open character offsets into text.
• value is the normalized full-unit value after applying sign, scale, and numeric format.
• dimensions records business segment, product, geography, investment, or another dimensional scope.
• evidence_url opens the corresponding number in the filing; source_url points to the SEC document.

XBRL labels are provided as reported by the filer and may reflect errors in the source filing.

Files and schema

Split	Passages	Companies	Facts	Typed-dimension facts
Train	4,922	4,113	87,539	4,181
Validation	776	765	14,360	375
Total	5,698	4,878	101,899	4,556

Train and validation are company-disjoint. The default files contain one nested row per passage; the flat/ directory provides equivalent passage, fact, and non-target tables.

Passage fields

Field	Description
chunk_id	Stable passage ID and flat-table join key
cik, ticker, company_name, sic_code	Issuer identity
accession, form_type, filed_at	Filing identity
fiscal_year, fiscal_period	Filing-level fiscal focus
chunk_type, heading, pre_context	Passage location and nearby context
text	Clean passage text; table cells are separated with vertical bars
source_url	SEC filing URL
facts_count	Number of target facts
scale_cue_in_text, sample_stratum	Presentation and sampling metadata
facts	Nested target-fact records
non_target_spans	Nested non-target numeric spans

Fact fields

Field	Description
fact_id	Stable fact ID
concept	Namespace-qualified XBRL concept
period	instant:DATE or duration:START..END
value	Precision-preserving normalized value
unit	Examples: monetary:USD, per_share:USD, shares, pure
displayed_text	Exact characters shown in the passage
text_start, text_end	Character offsets into text
scale	Power-of-ten presentation scale
decimals, format	Filing rounding and numeric-format metadata
dimensions	Explicit or typed dimensional scope
evidence_url	Link to the corresponding filing location

Each non_target_spans record contains span_text, start, end, and class.

Companion model

The companion model's training data includes these passages, so evaluating it on this dataset would be in-sample. Use separate held-out data for evaluation. The extraction prompt, output format, benchmark protocol, and model results are on the model card.

License and access

This sample is licensed under CC BY-NC 4.0. The license applies to StockAlloy's compilation, annotations, alignments, and derived metadata. The original SEC filings remain independently available through EDGAR.

Contact: data@stockalloy.com.

Evidence links are intended for inspection and light programmatic use. Contact StockAlloy before bulk retrieval.

• Dataset version: v1.2
• Selection manifest SHA-256: bc0b3e526742567daa5b17deacb533a4419e5cab4375962cdb5a0e0a7ef78a3a

@dataset{stockalloy_filedfact_passages_2026,
  title   = {FiledFact-Passages: Passage-Complete, Span-Grounded SEC Financial Facts},
  author  = {StockAlloy Research},
  year    = {2026},
  url     = {https://huggingface.co/datasets/StockAlloy/filedfact-passages}
}