README.md

metadata

license: cc-by-4.0
language:
  - en
pretty_name: EDGARCalcQA
task_categories:
  - question-answering
  - text-retrieval
tags:
  - financial-nlp
  - sec-filings
  - xbrl
  - numerical-reasoning
  - citation-grounding
  - retrieval-augmented-generation
size_categories:
  - 1K<n<10K

EDGARCalcQA Dataset Release

This folder contains the clean EDGARCalcQA release for research evaluation.

Contents

• benchmark_examples.json: JSON list of benchmark examples.
• benchmark_examples.jsonl: One benchmark example per line.
• benchmark_summary.txt: Human-readable benchmark summary.
• rejected_examples.json: Rejected candidates, if available.
• schema.json: Field names, JSON types, and short field descriptions.

Dataset Facts

• Companies: 60
• Filings: 1 recent 10-K per company
• Accepted examples: 513
• Rejected candidates: 12
• Task families: 222 value lookup, 138 derived ratio, 153 temporal comparison

Source Data

The benchmark uses public SEC 10-K filings and SEC XBRL companyfacts. Raw SEC filing HTML and processed chunk files are not bundled in this release folder by default.

Labels

Gold numeric answers are deterministic XBRL-backed labels. Value lookup answers use a single XBRL fact. Derived-ratio and temporal-comparison answers are recomputed from structured XBRL inputs. Gold evidence passages are filing-text chunks selected from the parsed 10-K text.

Evaluation

EDGARCalcQA supports strict and relaxed scoring. Strict scoring checks numeric tolerance and overlap with canonical gold evidence. Relaxed scoring accepts display-equivalent numeric forms and valid alternative support paths. Null or abstained answers count as failures.

Use

This dataset is for research evaluation of citation-grounded financial QA systems. It is not investment advice and should not be used for automated trading, credit or lending decisions, investment recommendations, compliance determinations, or other high-stakes financial decisions.