tayano/edinet-fraud-testset

EDINET fraud testset

A test set for accounting fraud detection on Japanese annual securities reports (有価証券報告書), released in the same format as EDINET-Bench (Sugiura et al., 2026).

EDINET-Bench undersamples the negative class, resulting in a positive rate of about 49%, which is far higher than the rate expected in practice. This test set does not undersample the negative class: it covers every annual securities report whose fiscal year end falls in a fixed period, and its positive rate is 0.11%. Evaluation subsets with different positive-class prevalences can be constructed by subsampling negatives, as done in the associated paper.

testset.parquet

Annual securities reports with a fiscal year end between 2016-03-31 and 2023-06-30. Of the 26,641 such reports, the 5,021 belonging to any of the 754 companies that appear in EDINET-Bench are excluded, leaving 21,620. The exclusion is applied at the company level.

	Count
Annual securities reports (FY ending 2016-03 to 2026-03)	36,940
Fiscal year end between 2016-03-31 and 2023-06-30	26,641
Reports of the 754 EDINET-Bench companies (excluded)	−5,021
Test set	21,620

• 21,620 reports, 3,287 companies
• 24 reports classified as accounting fraud (11 companies, 0.11%)

Column	Type	Description
doc_id	string	EDINET document ID
edinet_code	string	EDINET code of the filer
period_end	date32	Fiscal year end
text	string	Text of the report
label	int64	1 if we classified it as fraud in the light of ASCS 240

ASCS 240 is Auditing Standards Committee Statement No. 240, "The Auditor's Responsibilities Relating to Fraud in an Audit of Financial Statements" (監査基準報告書240「財務諸表監査における 不正」), issued by the Japanese Institute of Certified Public Accountants (JICPA). It distinguishes fraudulent misstatements from errors according to whether the misstatement is intentional, and applies to financial statement audits only.

bench_labels.parquet

For all 1,089 EDINET-Bench documents, our classification of errors and other corrections. This table does not contain the report text. EDINET-Bench is public, so it can be joined on doc_id.

EDINET-Bench uses the filing of an amended annual securities report as the starting point for positive-label assignment, and does not separate fraud from errors and other corrections. Following ASCS 240, we separate the two according to whether the misstatement is intentional. "Errors and other corrections" (誤謬等) covers errors as well as amendments touching only items outside 経理の状況.

Column	Type	Description
doc_id	string	EDINET document ID
is_error	int64	1 if we classified it as an error or other correction

Of the 534 documents that EDINET-Bench labels positive, 112 were classified as errors or other corrections.

How labels were assigned

1. Annual securities reports were downloaded from EDINET with edinet2dataset.
2. Each report was linked to the amended annual securities reports filed for the same company and the same fiscal year. Amendments are sometimes filed several years after the fiscal year ends, so the 6,863 amended reports filed up to 2026-06-30 were covered.
3. The stated reason for each amendment, taken from the first four pages of its PDF, was classified as fraud or as an error/other correction. We used the labeling procedure of edinet2dataset with three changes: the prompt was rewritten to follow ASCS 240 and to separate the two according to whether the misstatement is intentional; an amendment touching only items outside 経理の状況 was treated as an error or other correction; and a different large language model was used, because the one used by edinet2dataset is no longer available.
4. The classifications produced by the language model were reviewed by Tadashi Yano and cross-checked against daisanshaiinkai.com (third-party investigation reports) and the recommendations for administrative monetary penalty payment orders and the criminal referrals published by the Securities and Exchange Surveillance Commission.

How to use

from datasets import load_dataset
d = load_dataset("tayano/edinet-fraud-testset", "testset")
b = load_dataset("tayano/edinet-fraud-testset", "bench_labels")

Provenance and licensing

This dataset is licensed under the PDL 1.0 in accordance with EDINET's Terms of Use.

text is the body of annual securities reports published on EDINET. The documents were downloaded with edinet2dataset (Copyright 2025 SakanaAI, Apache License 2.0); the text was then extracted from the downloaded documents by Tadashi Yano. Please follow the terms of use of EDINET: https://disclosure2.edinet-fsa.go.jp/

Extraction and processing were carried out by Tadashi Yano (Yano Accounting Office). Neither the Financial Services Agency nor Sakana AI is responsible for the contents of this dataset.

Associated paper

Tadashi Yano and Yasuhiro Ogawa.
Re-examining Accounting Fraud Detection on EDINET-Bench:
Separating Error Labels and Evaluating under Class Imbalance.
JSAI Technical Report, Type 2 SIG, The 37th SIG-FIN, October 2026.

EDINET-Bench:

@inproceedings{
sugiura2026edinetbench,
title={{EDINET}-Bench: Evaluating {LLM}s on Complex Financial Tasks using Japanese Financial Statements},
author={Issa Sugiura and Takashi Ishida and Taro Makino and Chieko Tazuke and Takanori Nakagawa and Kosuke Nakago and David Ha},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=Dxns0cj15A}
}

---

日本語の説明は README_ja.md にあります。