Datasets:

EDGARCalcQA
/

edgarcalcqa

	{
	"@context": {
	"@language": "en",
	"@vocab": "https://schema.org/",
	"cr": "http://mlcommons.org/croissant/",
	"rai": "http://mlcommons.org/croissant/RAI/",
	"sc": "https://schema.org/",
	"prov": "http://www.w3.org/ns/prov#"
	},
	"@type": "sc:Dataset",
	"name": "EDGARCalcQA",
	"description": "EDGARCalcQA is an XBRL-grounded benchmark for citation-auditable numerical reasoning over SEC 10-K filings.",
	"url": "https://huggingface.co/datasets/EDGARCalcQA/edgarcalcqa",
	"codeRepository": "https://github.com/edgarcalcqa/EDGARCalcQA",
	"license": "https://creativecommons.org/licenses/by/4.0/",
	"creator": [
	{
	"@type": "Person",
	"name": "Anonymous Authors"
	}
	],
	"citation": "Anonymous Authors. EDGARCalcQA: An XBRL-Grounded Benchmark for Citation-Auditable Numerical Reasoning over SEC Filings. 2026.",
	"keywords": [
	"financial question answering",
	"SEC filings",
	"XBRL",
	"citation grounding",
	"numerical reasoning",
	"retrieval augmented generation"
	],
	"datePublished": "2026-05-05",
	"inLanguage": "en",
	"isAccessibleForFree": true,
	"licenseNotes": "CC BY 4.0 applies to EDGARCalcQA annotations and metadata. Original SEC filings and SEC XBRL companyfacts remain public SEC source material.",
	"dataSource": [
	"Public SEC EDGAR 10-K filings",
	"SEC XBRL companyfacts"
	],
	"collectionProcess": "The release is built from public SEC EDGAR 10-K filings and SEC XBRL companyfacts. The pipeline downloads one recent 10-K per selected public company, parses filing text into sectioned chunks, and reads XBRL facts from SEC companyfacts records. Raw SEC filings are not redistributed in the released annotation package.",
	"annotationProcess": "Questions and numeric labels are generated deterministically from supported US-GAAP XBRL concepts. Value lookup examples use a single annual fact; derived ratio and temporal comparison examples are recomputed from structured XBRL inputs. Gold evidence passages are selected from filing-text chunks and validated with schema, arithmetic, period-alignment, numeric-form, and evidence-support checks.",
	"datasetComposition": {
	"examples": 513,
	"companies": 60,
	"rejectedCandidates": 12,
	"questionTypes": {
	"value_lookup": 222,
	"derived_ratio": 138,
	"temporal_comparison": 153
	}
	},
	"intendedUse": [
	"Research evaluation of financial numerical reasoning",
	"Citation grounding",
	"Retrieval-augmented generation over long SEC filings",
	"Long-context model evaluation"
	],
	"outOfScopeUse": [
	"Investment advice",
	"Automated trading",
	"Automated credit/lending decisions",
	"High-stakes financial decisions"
	],
	"limitations": [
	"Public-company bias",
	"10-K focus",
	"Standardized XBRL concept focus",
	"Limited footnotes and segment disclosures",
	"English-language SEC filing setting",
	"Evidence is chunk-based and may not capture every valid support path"
	],
	"sensitiveData": {
	"description": "Sensitive personal data is not intentionally collected. Source documents are public corporate disclosures.",
	"containsSensitivePersonalInformation": false
	},
	"maintenanceAndVersioning": "Dataset releases are tied to fixed annotation files, cached prediction files, evaluation reports, and Croissant metadata. Changes to company coverage, example counts, scoring, or annotation schema should be documented in release notes and should trigger regeneration of cached baseline outputs.",
	"distribution": [
	{
	"@type": "sc:FileObject",
	"@id": "benchmark_examples_json",
	"name": "benchmark_examples_json",
	"description": "JSON list of EDGARCalcQA benchmark examples.",
	"contentUrl": "release/dataset/benchmark_examples.json",
	"encodingFormat": "application/json",
	"sha256": "fcb7965474902462bd0b37deaf6e585eef8074afd616df8903cab460a2c1a6bf"
	},
	{
	"@type": "sc:FileObject",
	"@id": "benchmark_examples_jsonl",
	"name": "benchmark_examples_jsonl",
	"description": "Line-delimited benchmark examples.",
	"contentUrl": "release/dataset/benchmark_examples.jsonl",
	"encodingFormat": "application/jsonl",
	"sha256": "cc0a3e8fbf113caf11ec0cc6a2357d9651a80829d5c970e43d6e829401d87a0c"
	},
	{
	"@type": "sc:FileObject",
	"@id": "benchmark_summary",
	"name": "benchmark_summary",
	"description": "Human-readable benchmark summary.",
	"contentUrl": "release/dataset/benchmark_summary.txt",
	"encodingFormat": "text/plain",
	"sha256": "f8206f9cd91b50f35c741990f6164bc450b7bbef6991213a75d7602a678cf937"
	},
	{
	"@type": "sc:FileObject",
	"@id": "schema_json",
	"name": "schema_json",
	"description": "Field schema summary for benchmark examples.",
	"contentUrl": "release/dataset/schema.json",
	"encodingFormat": "application/json",
	"sha256": "ad923cebaa697a56b47d3dacc02ee016585dadb5f3388b6df430c9c9752a77a0"
	},
	{
	"@type": "sc:FileObject",
	"@id": "dataset_readme",
	"name": "dataset_readme",
	"description": "Release dataset README.",
	"contentUrl": "release/dataset/README.md",
	"encodingFormat": "text/markdown",
	"sha256": "2fcc74730015f824b6fa2db1f474cb572f4df74b3f83bef4e8c20ca40981e228"
	},
	{
	"@type": "sc:FileObject",
	"@id": "rejected_examples_json",
	"name": "rejected_examples_json",
	"description": "Rejected candidate examples and rejection reasons.",
	"contentUrl": "release/dataset/rejected_examples.json",
	"encodingFormat": "application/json",
	"sha256": "27190a6662c25d11c1e247895943976f2c5df3162b47ace50b160871d769c977"
	}
	],
	"recordSet": [
	{
	"@type": "cr:RecordSet",
	"@id": "examples",
	"name": "benchmark_examples",
	"description": "EDGARCalcQA benchmark examples with XBRL-backed labels and evidence metadata.",
	"field": [
	{
	"@type": "cr:Field",
	"@id": "examples/id",
	"name": "id",
	"description": "Stable benchmark example identifier.",
	"dataType": "sc:Text",
	"source": {
	"fileObject": {
	"@id": "benchmark_examples_jsonl"
	},
	"extract": {
	"jsonPath": "$.id"
	}
	}
	},
	{
	"@type": "cr:Field",
	"@id": "examples/company",
	"name": "company",
	"description": "Company name.",
	"dataType": "sc:Text",
	"source": {
	"fileObject": {
	"@id": "benchmark_examples_jsonl"
	},
	"extract": {
	"jsonPath": "$.company"
	}
	}
	},
	{
	"@type": "cr:Field",
	"@id": "examples/ticker",
	"name": "ticker",
	"description": "Company ticker symbol.",
	"dataType": "sc:Text",
	"source": {
	"fileObject": {
	"@id": "benchmark_examples_jsonl"
	},
	"extract": {
	"jsonPath": "$.ticker"
	}
	}
	},
	{
	"@type": "cr:Field",
	"@id": "examples/question_type",
	"name": "question_type",
	"description": "Task family.",
	"dataType": "sc:Text",
	"source": {
	"fileObject": {
	"@id": "benchmark_examples_jsonl"
	},
	"extract": {
	"jsonPath": "$.question_type"
	}
	}
	},
	{
	"@type": "cr:Field",
	"@id": "examples/question",
	"name": "question",
	"description": "Natural-language question.",
	"dataType": "sc:Text",
	"source": {
	"fileObject": {
	"@id": "benchmark_examples_jsonl"
	},
	"extract": {
	"jsonPath": "$.question"
	}
	}
	},
	{
	"@type": "cr:Field",
	"@id": "examples/answer_numeric",
	"name": "answer_numeric",
	"description": "Gold numeric answer.",
	"dataType": "sc:Float",
	"source": {
	"fileObject": {
	"@id": "benchmark_examples_jsonl"
	},
	"extract": {
	"jsonPath": "$.answer_numeric"
	}
	}
	},
	{
	"@type": "cr:Field",
	"@id": "examples/answer_display",
	"name": "answer_display",
	"description": "Human-readable gold answer.",
	"dataType": "sc:Text",
	"source": {
	"fileObject": {
	"@id": "benchmark_examples_jsonl"
	},
	"extract": {
	"jsonPath": "$.answer_display"
	}
	}
	},
	{
	"@type": "cr:Field",
	"@id": "examples/unit",
	"name": "unit",
	"description": "Answer unit or scale.",
	"dataType": "sc:Text",
	"source": {
	"fileObject": {
	"@id": "benchmark_examples_jsonl"
	},
	"extract": {
	"jsonPath": "$.unit"
	}
	}
	},
	{
	"@type": "cr:Field",
	"@id": "examples/gold_concepts",
	"name": "gold_concepts",
	"description": "XBRL concepts used for the label.",
	"dataType": "sc:Text",
	"source": {
	"fileObject": {
	"@id": "benchmark_examples_jsonl"
	},
	"extract": {
	"jsonPath": "$.gold_concepts"
	}
	}
	},
	{
	"@type": "cr:Field",
	"@id": "examples/gold_period",
	"name": "gold_period",
	"description": "Fiscal period metadata.",
	"dataType": "sc:Text",
	"source": {
	"fileObject": {
	"@id": "benchmark_examples_jsonl"
	},
	"extract": {
	"jsonPath": "$.gold_period"
	}
	}
	},
	{
	"@type": "cr:Field",
	"@id": "examples/form",
	"name": "form",
	"description": "SEC filing form type.",
	"dataType": "sc:Text",
	"source": {
	"fileObject": {
	"@id": "benchmark_examples_jsonl"
	},
	"extract": {
	"jsonPath": "$.form"
	}
	}
	},
	{
	"@type": "cr:Field",
	"@id": "examples/filed",
	"name": "filed",
	"description": "SEC filing date.",
	"dataType": "sc:Date",
	"source": {
	"fileObject": {
	"@id": "benchmark_examples_jsonl"
	},
	"extract": {
	"jsonPath": "$.filed"
	}
	}
	},
	{
	"@type": "cr:Field",
	"@id": "examples/derivation",
	"name": "derivation",
	"description": "Structured derivation metadata.",
	"dataType": "sc:Text",
	"source": {
	"fileObject": {
	"@id": "benchmark_examples_jsonl"
	},
	"extract": {
	"jsonPath": "$.derivation"
	}
	}
	},
	{
	"@type": "cr:Field",
	"@id": "examples/gold_evidence",
	"name": "gold_evidence",
	"description": "Gold evidence passage metadata.",
	"dataType": "sc:Text",
	"source": {
	"fileObject": {
	"@id": "benchmark_examples_jsonl"
	},
	"extract": {
	"jsonPath": "$.gold_evidence"
	}
	}
	},
	{
	"@type": "cr:Field",
	"@id": "examples/gold_exact",
	"name": "gold_exact",
	"description": "Canonical exact gold value metadata.",
	"dataType": "sc:Text",
	"source": {
	"fileObject": {
	"@id": "benchmark_examples_jsonl"
	},
	"extract": {
	"jsonPath": "$.gold_exact"
	}
	}
	},
	{
	"@type": "cr:Field",
	"@id": "examples/accepted_numeric_forms",
	"name": "accepted_numeric_forms",
	"description": "Accepted numeric forms for evaluation.",
	"dataType": "sc:Text",
	"source": {
	"fileObject": {
	"@id": "benchmark_examples_jsonl"
	},
	"extract": {
	"jsonPath": "$.accepted_numeric_forms"
	}
	}
	},
	{
	"@type": "cr:Field",
	"@id": "examples/support_units",
	"name": "support_units",
	"description": "Support units for relaxed citation scoring.",
	"dataType": "sc:Text",
	"source": {
	"fileObject": {
	"@id": "benchmark_examples_jsonl"
	},
	"extract": {
	"jsonPath": "$.support_units"
	}
	}
	},
	{
	"@type": "cr:Field",
	"@id": "examples/accepted_support_paths",
	"name": "accepted_support_paths",
	"description": "Accepted relaxed citation support paths.",
	"dataType": "sc:Text",
	"source": {
	"fileObject": {
	"@id": "benchmark_examples_jsonl"
	},
	"extract": {
	"jsonPath": "$.accepted_support_paths"
	}
	}
	}
	]
	}
	],
	"rai:dataLimitations": "EDGARCalcQA is limited to public SEC 10-K annual reports and standardized XBRL companyfacts for U.S.-listed public companies. It focuses on numerical reasoning over common financial-statement concepts such as revenue, net income, operating cash flow, total assets, derived margins, and year-over-year changes. The dataset does not cover all filing types, private companies, non-U.S. reporting regimes, non-English disclosures, footnotes in full detail, segment-level disclosures, or highly specialized financial concepts. Evidence is chunk-based, so some valid support paths may be distributed across multiple passages or appear in alternative filing locations not selected as the gold evidence. The dataset is intended for research evaluation of numerical reasoning, retrieval, and citation grounding. It is not recommended for investment advice, automated trading, credit/lending decisions, regulatory determinations, or other high-stakes financial decisions.",
	"rai:dataBiases": "The dataset has selection bias toward large, stable, U.S.-listed public companies with accessible SEC 10-K filings and relatively standardized XBRL reporting. It under-represents smaller companies, private firms, financial institutions with specialized reporting, foreign private issuers, non-U.S. accounting standards, and companies with unusual or highly complex filings. Labels are derived from XBRL facts, so the benchmark favors concepts that are consistently tagged in machine-readable form and may under-represent nuanced disclosures found only in footnotes, MD&A narratives, or segment tables. Models evaluated on this dataset may appear stronger on standardized financial-statement extraction than on broader financial analysis, qualitative reasoning, or non-standard reporting contexts.",
	"rai:personalSensitiveInformation": "The dataset does not intentionally collect personal or sensitive information about individuals. The source documents are public corporate SEC filings and machine-readable XBRL companyfacts. The dataset represents corporate financial disclosures rather than individual-level data. It may contain names of company executives, directors, auditors, or other public corporate officers if those names appear in public SEC filings, but these are not the target of the benchmark and are not used as labels. The dataset does not intentionally include gender, age, health or medical data, political or religious beliefs, socio-economic status of individuals, or other sensitive personal attributes.",
	"rai:dataUseCases": "EDGARCalcQA is intended to measure citation-auditable numerical reasoning over SEC 10-K filings. It evaluates whether a system can answer financial numerical questions accurately and provide evidence from the filing text that supports the answer. Validated use cases include research evaluation of financial question answering, retrieval-augmented generation over long financial documents, long-context model evaluation, numeric answer extraction, derived financial-ratio computation, temporal comparison, and citation/evidence grounding. The dataset has not been validated for investment recommendations, trading strategies, credit decisions, legal compliance determinations, audit opinions, or production financial advisory systems.",
	"rai:dataSocialImpact": "Potential positive impacts include improving the reliability, transparency, and auditability of AI systems used for financial document analysis. By requiring both numeric correctness and evidence grounding, the dataset encourages systems that provide verifiable answers rather than unsupported financial claims. Potential negative impacts include misuse as a proxy for real investment analysis, overconfidence in models evaluated only on standardized 10-K numerical questions, or deployment in high-stakes financial workflows without human review. Mitigations include clear documentation that the dataset is for research evaluation only, explicit out-of-scope use restrictions, citation-based evaluation, and release of benchmark metadata and scoring code to support reproducibility and error analysis.",
	"rai:hasSyntheticData": false,
	"prov:wasDerivedFrom": [
	{
	"@id": "https://www.sec.gov/edgar/search/",
	"prov:label": "SEC EDGAR 10-K filings",
	"sc:license": "Public domain / U.S. government public data",
	"prov:wasAttributedTo": {
	"@id": "https://www.sec.gov/",
	"prov:label": "U.S. Securities and Exchange Commission"
	}
	},
	{
	"@id": "https://data.sec.gov/",
	"prov:label": "SEC XBRL companyfacts API",
	"sc:license": "Public domain / U.S. government public data",
	"prov:wasAttributedTo": {
	"@id": "https://www.sec.gov/",
	"prov:label": "U.S. Securities and Exchange Commission"
	}
	}
	],
	"prov:wasGeneratedBy": [
	{
	"@type": "prov:Activity",
	"prov:type": {
	"@id": "https://www.wikidata.org/wiki/Q4929239"
	},
	"prov:label": "SEC filing and XBRL fact collection",
	"sc:description": "Collected public SEC 10-K filing HTML documents and SEC XBRL companyfacts for selected U.S.-listed public companies. The company universe was selected to cover multiple sectors while prioritizing companies with accessible annual filings and standardized XBRL facts. The collection uses public SEC EDGAR sources and does not involve private, proprietary, or user-submitted data.",
	"prov:wasAttributedTo": [
	{
	"@type": "prov:Agent",
	"@id": "anonymous_authors",
	"prov:label": "Anonymous Authors",
	"sc:description": "Human research team responsible for selecting the company universe and running the collection pipeline. Institutional affiliation, geographic location, and identifying details are withheld during double-blind review."
	}
	]
	},
	{
	"@type": "prov:Activity",
	"prov:type": {
	"@id": "https://www.wikidata.org/wiki/Q5227332"
	},
	"prov:label": "Filing parsing and XBRL filtering",
	"sc:description": "Automated Python scripts parsed filing HTML into sections and overlapping text chunks. XBRL companyfacts were filtered to annual fiscal-year observations and mapped to common financial concepts such as revenue, net income, operating cash flow, total assets, gross margin inputs, net margin inputs, and year-over-year comparison values. Candidate records with missing facts, invalid periods, duplicate identifiers, or unusable evidence were rejected.",
	"prov:wasAttributedTo": [
	{
	"@type": "prov:SoftwareAgent",
	"@id": "automated_edgarcalcqa_preprocessing_pipeline",
	"prov:label": "Automated EDGARCalcQA preprocessing pipeline",
	"sc:description": "Synthetic/software agent consisting of deterministic Python scripts for HTML parsing, chunking, XBRL filtering, concept mapping, and candidate rejection. No LLM or generative model was used for preprocessing."
	}
	]
	},
	{
	"@type": "prov:Activity",
	"prov:type": {
	"@id": "https://www.wikidata.org/wiki/Q109719325"
	},
	"prov:label": "XBRL-grounded label and evidence construction",
	"sc:description": "Benchmark labels were generated deterministically from SEC XBRL facts. Value-lookup answers use single XBRL facts, derived-ratio answers compute ratios from two XBRL facts, and temporal-comparison answers compute year-over-year percentage changes from current and prior fiscal-year facts. Questions were template-generated from task family, company, fiscal period, and target concept. Gold evidence passages were selected from parsed filing chunks using concept, section, keyword, and numeric-match heuristics. No crowdworkers or external human annotators were used.",
	"prov:wasAttributedTo": [
	{
	"@type": "prov:SoftwareAgent",
	"@id": "automated_edgarcalcqa_annotation_pipeline",
	"prov:label": "Automated EDGARCalcQA annotation pipeline",
	"sc:description": "Synthetic/software agent consisting of deterministic Python scripts for question templating, XBRL-backed answer generation, derivation construction, accepted numeric form generation, and heuristic evidence selection. No LLM was used to generate gold answers, gold labels, or gold evidence."
	}
	]
	},
	{
	"@type": "prov:Activity",
	"prov:type": {
	"@id": "https://www.wikidata.org/wiki/Q3306762"
	},
	"prov:label": "Targeted evidence audit and automated validation",
	"sc:description": "Validated the benchmark using automated checks for unique IDs, complete schema fields, arithmetic recomputation, fiscal period alignment, accepted numeric forms, non-null gold evidence, and rejected-example tracking. Evidence quality was additionally reviewed through targeted audits of sampled examples and warning cases, especially examples with missing year cues or missing detected values. Examples with weak or unsupported evidence were fixed or removed before release.",
	"prov:wasAttributedTo": [
	{
	"@type": "prov:Agent",
	"@id": "anonymous_authors",
	"prov:label": "Anonymous Authors",
	"sc:description": "Human research team responsible for targeted evidence review, manual inspection of warning cases, and final inclusion/removal decisions. Institutional affiliation, geographic location, and identifying details are withheld during double-blind review. No crowdsourcing was used."
	},
	{
	"@type": "prov:SoftwareAgent",
	"@id": "automated_edgarcalcqa_validation_pipeline",
	"prov:label": "Automated EDGARCalcQA validation pipeline",
	"sc:description": "Synthetic/software agent consisting of deterministic Python scripts for schema validation, arithmetic recomputation, fiscal-period checking, evidence non-null checks, duplicate detection, and rejected-example tracking. No LLM was used for validation decisions."
	}
	]
	}
	],
	"version": "1.0"
	}