jfinqa: Japanese Financial Numerical Reasoning QA Benchmark

jfinqa is a benchmark for evaluating LLMs on Japanese financial question answering that requires numerical reasoning over real corporate disclosures.

Unlike existing Japanese financial benchmarks that focus on classification or simple lookup, jfinqa requires models to cross-reference text and tables to perform multi-step calculations.

Dataset Summary

	Count
Total questions	1000
Companies	68
Accounting standards	J-GAAP (582), IFRS (377), US-GAAP (41)
Avg. program steps	2.59
Avg. table rows	13.3
Max program steps	6 (DuPont decomposition)

Subtasks

Config	Description	Count
numerical_reasoning	Calculate financial metrics from table data	550
consistency_checking	Verify internal consistency of reported figures	200
temporal_reasoning	Analyze trends and changes across periods	250

Usage

from datasets import load_dataset

# Load all questions
ds = load_dataset("ajtgjmdjp/jfinqa", "all", split="test")

# Load a specific subtask
ds_nr = load_dataset("ajtgjmdjp/jfinqa", "numerical_reasoning", split="test")

With the jfinqa library

from jfinqa import load_dataset, evaluate

questions = load_dataset("numerical_reasoning")
result = evaluate(questions, predictions={"nr_001": "25.0%"})
print(result.summary())

Install: pip install jfinqa

Data Format

Each example contains:

Field	Type	Description
id	string	Unique identifier (e.g., nr_001, cc_015, tr_042)
subtask	string	Task category
company_name	string	Company name in Japanese
edinet_code	string	EDINET company code
source_doc_id	string	Source document ID for provenance
filing_year	string	Fiscal year
accounting_standard	string	J-GAAP, IFRS, or US-GAAP
scale	string	Number unit (百万円, 億円, etc.)
pre_text	list[string]	Context paragraphs before the table
post_text	list[string]	Context paragraphs after the table
table_headers	list[string]	Table column headers
table_rows	list[list[string]]	Table data rows
question	string	Question in Japanese
answer	string	Gold answer
program	list[string]	FinQA-compatible DSL program
gold_evidence	list[int]	Indices of relevant table rows

The data format is compatible with FinQA (Chen et al., 2021).

Example

{
  "id": "tr_010",
  "subtask": "temporal_reasoning",
  "company_name": "味の素",
  "accounting_standard": "IFRS",
  "question": "味の素は2023年3月期から2024年3月期にかけて増収か減収か。",
  "answer": "増収",
  "program": ["subtract(1439231.0, 1359115.0)", "greater(#0, 0)"],
  "table_headers": ["", "2024年3月期", "2023年3月期"],
  "table_rows": [
    ["売上高", "1,439,231", "1,359,115"],
    ["売上原価", "927,783", "888,727"],
    ["売上総利益", "511,448", "470,387"]
  ]
}

Values are reported in million yen (百万円) and correspond to the company's consolidated filings retrieved from EDINET.

Japanese-Specific Features

• Accounting standards: Handles J-GAAP (経常利益), IFRS, and US-GAAP differences
• Number formats: △ for negative, 百万円/億円 units, full-width digits
• Japanese financial terminology: 売上高, 営業利益, 経常利益, etc.

Data Source

Financial data is obtained from EDINET (Electronic Disclosure for Investors' NETwork), operated by the Financial Services Agency of Japan (金融庁). EDINET data is provided under the Public Data License 1.0.

Citation

@misc{jfinqa2025,
  title={jfinqa: Japanese Financial Numerical Reasoning QA Benchmark},
  author={ajtgjmdjp},
  year={2025},
  url={https://github.com/ajtgjmdjp/jfinqa},
}

Related

• jfinqa (GitHub) — Evaluation library and CLI
• edinet-mcp — EDINET XBRL parser (companion project)
• FinQA — English financial QA benchmark
• EDINET-Bench — Sakana AI's financial classification benchmark

License

Apache-2.0. See NOTICE for third-party attributions.