jrwana/sec-filing-rag-eval

SEC 10-K XBRL-Derived Retrieval Eval

A retrieval / grounded-QA benchmark over eight SEC 10-K filings (AAPL, MSFT, NVDA, AMD; FY2023–24). The distinguishing feature is that every gold label is derived mechanically from the filing's inline XBRL — not hand-annotated. No financial-domain judgement enters the labels, so the ground truth is auditable and reproducible: each answer is a specific tagged fact, and each question's gold evidence is the set of chunks that contain the XBRL occurrence of that fact.

• Code, pipeline, and retrieval benchmark: https://github.com/jrwana/sec-filing-rag
• Live demo (retrieval + grounded generation): https://jrwana-sec-filing-rag.hf.space/

What's here

Config	Rows	What each row is
answerable	491	A question about one XBRL fact, with its canonical value, unit, value type, and the gold evidence anchors.
unanswerable	98	A near-miss: names a real metric for the company but in a period the filing does not report. The correct behaviour is to abstain.

Both configs share a held-out-by-company split — there is no company overlap between dev and test, so tuning on dev cannot leak company-specific phrasing into the reported test numbers:

• dev — AAPL, MSFT
• test — NVDA, AMD

How the gold is derived (the exhibit)

1. Extract every inline-XBRL fact from the filing (us-gaap:* / company concepts) with its context: period, unit, and dimensional axes.
2. Canonicalise each fact to a stable key — COMPANY|concept|period|unit|dimensions — so the same economic quantity has one identity across filings and restatements.
3. Template a question from the concept label and context, then paraphrase it (answerable rows carry both question and question_paraphrased; the benchmark scores the paraphrase).
4. Gold evidence = the chunks whose text spans the fact's XBRL occurrence. anchors lists the fact occurrences; the repo's gold map resolves those to chunk ids for retrieval scoring.
5. Near-misses are constructed by taking a real metric and shifting it to a period the filing does not cover (meta.absent_period_fy), verified against the periods actually present (meta.checked_against_inventory).

Because step 4 is mechanical, "did retrieval surface the gold evidence?" and "is the generated number correct?" are both exact, tolerance-typed checks rather than judgement calls.

Field schema

answerable

Field	Type	Meaning
question_id	str	Stable id (q…).
question	str	Templated question.
question_paraphrased	str	Natural-language paraphrase — this is what the benchmark asks.
company	str	AAPL / MSFT / NVDA / AMD.
split	str	dev or test (by company, see above).
canonical_fact_key	str	COMPANY|concept|period|unit|dimensions — the fact's identity.
canonical_value	str	Gold numeric value in base units (string to preserve precision).
unit	str	e.g. USD, USD/shares, shares, pure.
value_type	str	monetary / count / per_share / percentage — sets the numeric tolerance.
output_representation	str	How the value is expressed (e.g. base_units).
anchors	list[str]	Gold fact occurrences (FILING|f-id); resolve to chunks via the repo's gold map.
scope	str	fact (a single tagged value) or filing.
category	str	dimensional / temporal / direct.
label_source	str	visible (rendered in the filing) or taxonomy.
meta	struct	occurrences, restated, dimensions, period.

unanswerable — same envelope; canonical_value / unit / value_type / output_representation are null and anchors is empty. meta carries absent_period_fy (the period the question asks about) and checked_against_inventory (the periods the filing actually reports).

Distribution (491 answerable)

• By company: AAPL 122 AMD 120 MSFT 120 NVDA 129
• By value type: monetary 456 count 21 percentage 10 per_share 4
• By category: dimensional 240 temporal 240 direct 11

Loading

from datasets import load_dataset

ans = load_dataset("jrwana/sec-filing-rag-eval", "answerable")          # dev + test splits
nm  = load_dataset("jrwana/sec-filing-rag-eval", "unanswerable", split="test")

row = ans["test"][0]
row["question_paraphrased"], row["canonical_value"], row["anchors"]

The four JSONL files are the exact bytes the benchmark scores against, so any number reported in the repo can be reproduced against this dataset.

Provenance & license

• Underlying facts come from public filings on SEC EDGAR (inline XBRL). The dataset stores derived numeric facts and generated questions, not filing prose.
• Questions and gold labels (this dataset): CC-BY-4.0.
• The companies named are the subjects of their own public filings; nothing here is investment advice.

Caveats

• Retrieval is the ceiling, not the generator. On the held-out test split, gold evidence reaches the top-5 for only ~44% of answerable questions, which bounds end-to-end accuracy regardless of the model. See the repo for the full breakdown.
• Labels reflect what is XBRL-tagged; a figure discussed only in prose and never tagged is out of scope by construction.
• Eight filings, four companies, two fiscal years — the set is designed for auditable methodology, not broad coverage.