XBRLFiling

600 financial calculation questions · 28 public companies · 86 annual filings

XBRLFiling evaluates numerical question answering over public-company financial statements. Each question asks for a financial subtotal whose target value has been removed from the supplied statement. The remaining line items provide the evidence needed to calculate the answer. A filing-declared XBRL calculation relationship supplies the reference calculation and provenance.

This release preserves the existing benchmark questions, evidence, answers, and numeric values. Workspace metadata containing author or laboratory identifiers and local filesystem paths is excluded; public company names, tickers, SEC identifiers, and source links remain available for traceability.

Quick start

Load the test split through the anonymous review link. No Hugging Face account or token is required.

from datasets import load_dataset

dataset = load_dataset(
    "parquet",
    data_files={
        "test": "https://anonymous-hf.com/api/a/dc78cftylin7/resolve/data/test.parquet"
    },
    split="test",
    token=False,
)
example = dataset[0]

# Give the answer generator only the question and the redacted evidence.
model_input = {
    "question": example["question"],
    "context": example["context"],
}

# Keep the reference answer and proof separate for evaluation.
reference_answer = example["answer"]

The single test split contains all 600 benchmark examples. No training or validation split is introduced. The dataset card explicitly selects the Parquet file, so the equivalent JSONL copy does not create duplicate examples when loaded with datasets.

Coverage

Statement family	Target concepts	Questions
Income statement	5	226
Balance sheet	7	279
Cash flow statement	3	95
Total	15	600

The frozen source records have fiscal-year labels from 2021 through 2026. The unit is USD millions. Each question specifies its requested rounding. There are 2,490 reference operands, with 2–16 operands per calculation.

Fields

One row represents one question and its reference answer.

Field	Description
id	Stable benchmark question identifier.
question	Original natural-language question.
context	Rendered financial statement with the target cell redacted.
answer	Original formatted reference answer.
answer_value	Numeric reference value, in answer_unit.
answer_unit	Unit of the reference value: USD millions.
ticker	Public-company ticker symbol.
cik	SEC Central Index Key.
filing_id	Stable filing label from the source benchmark.
fiscal_year	Fiscal-year label associated with the question.
period_end	Source financial period end.
report_date	Report-date field preserved from the source metadata.
statement_family	Income statement, balance sheet, or cash flow family.
target_concept	XBRL concept requested by the question.
calculation	Reference additive relationship declared by the filing.
operands	Reference operands, including concept, weight, and value in USD millions.
source_url	Public SEC source-filing URL.
accession	SEC filing accession number.
proof_source	xbrl_calculation_linkbase for every example.

Evaluation boundary: answer, answer_value, calculation, and operands are reference labels or proof material. Do not include them in the answer generator's prompt. Use question and context for the provided-evidence task; report retrieval-based experiments separately.

Files

.
├── README.md
├── LICENSE.md
├── NOTICE.md
├── data/
│   ├── test.parquet
│   └── test.jsonl
└── metadata/
    ├── filings.jsonl
    ├── quality_report.json
    └── manifest.json

test.parquet and test.jsonl contain the same 600 rows. The filing index contains the 86 source filings and public URLs. The quality report records release validation checks, and the manifest records SHA-256 file checksums.

This compact release contains the evidence needed for the provided-evidence benchmark. Raw XBRL XML, SEC companyfacts snapshots, retrieval corpora, model weights, and experiment outputs are not included. Reproducing the original fact-binding or retrieval pipeline requires those additional source artifacts.

Construction and sources

The source benchmark selected additive relationships from Form 10-K XBRL calculation linkbases. It resolved the parent and child facts for the filing period, checked the calculation against a construction tolerance, rendered a multi-period statement, and redacted the target cell. Deterministic round-robin selection across companies produced the 600-question benchmark.

The financial facts originate in issuer filings available through SEC EDGAR. Each record retains its filing accession and source URL. Consult the SEC developer resources before automatically downloading source filings.

Quality and limitations

• The source audit found 600 unique question identifiers and question strings, one evidence block and target redaction per row, and no missing answers or required provenance fields. Release checks are recorded in metadata/quality_report.json.
• Stored reference operands and values use three decimal places in USD millions. In 29 rows, recomputing the weighted sum from the stored operands leaves a nonzero residual; the maximum absolute residual is USD 2 million. All rows satisfy the original construction tolerance: the larger of USD 1 million and 0.5% of the absolute parent value. These residuals are not attributed solely to decimal rounding, and exact equality of every stored operand sum is not guaranteed.
• The benchmark covers additive XBRL relationships and 28 selected issuers; it is not representative of all financial reasoning or all public companies.
• Evidence is rendered from XBRL rather than taken directly from a complete filing or PDF. A provided-evidence evaluation does not measure retrieval.
• Filings, taxonomies, and restatements may change. This release describes the frozen benchmark; rebuilding from newer EDGAR data can yield different rows.

License and attribution

The dataset compilation and original contributions are released under CC BY-NC 4.0. Credit XBRLFiling contributors, link to this dataset and the license, and indicate changes when sharing an adaptation. See LICENSE.md for the complete terms and NOTICE.md for source attribution and scope.

Public-company filings and their identifiers retain their source attribution. The release license grants only rights held by the dataset contributors; it does not relicense third-party material or imply SEC or issuer endorsement.