You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
By requesting access to this dataset, you agree to use it responsibly and not for harmful, unlawful, or unethical purposes.
Log in or Sign Up to review the conditions and access this dataset content.
APTO-SorryBench-JA
APTO-SorryBench-JA is a Japanese translated and annotated version of the original Sorry-Bench dataset for AI safety evaluation research.
Sorry-Bench is a benchmark designed to evaluate whether Large Language Models (LLMs) appropriately refuse harmful requests while still providing helpful responses to safe requests.
The original English prompts are preserved alongside the Japanese translations to improve traceability and facilitate comparison with the original dataset.
This dataset includes:
- Japanese translations of the original Sorry-Bench prompts
- Japanese reference responses
- Refusal reasoning annotations
- Additional safety annotations created by APTO, Inc.
Creation Process
Translation and Localization
The original Sorry-Bench prompts were translated into natural Japanese by APTO, Inc.
Translations were carefully localized to preserve the original evaluation intent while ensuring that the prompts sound natural for Japanese speakers.
Modifications
Compared with the original Sorry-Bench dataset, APTO has added:
- Japanese translations
- Japanese reference responses
- Refusal reasoning annotations
- Additional safety annotations (
risk_type,intent,abuse_type,refusal_required)
The original English prompts and categories are preserved for reference.
Reference Response Annotation
Reference responses were manually created or reviewed in Japanese to support evaluation of model refusal behavior.
The responses are intended to evaluate whether models:
- Appropriately refuse harmful requests
- Avoid unnecessary refusals
- Provide safe alternative assistance whenever appropriate
Refusal Reasoning Annotation
Each sample includes a manually written explanation describing why the request should be refused (or answered).
These annotations are intended to support AI safety research, explainability, and LLM-as-a-Judge evaluation.
Additional Safety Annotation
Each sample contains additional metadata created by APTO:
risk_typeintentabuse_typerefusal_required
These annotations were assigned based on APTO's internal annotation guidelines and manually reviewed to improve consistency across the dataset.
Intended Use
This dataset is intended for:
- AI safety research
- Benchmark evaluation
- Model evaluation
- LLM-as-a-Judge
- Safety fine-tuning
- Alignment research
- Japanese AI safety evaluation
- Commercial or internal safety evaluation
This dataset must not be used to facilitate harmful, unlawful, or unethical activities.
License
This dataset is based on the original Sorry-Bench dataset.
The Japanese translation and additional annotations were created by APTO, Inc.
Before publishing this Japanese version, APTO contacted the original authors directly and received explicit permission to release this translated version.
Users should comply with the original Sorry-Bench license and terms of use.
Acknowledgements
This dataset is based on the original Sorry-Bench benchmark.
Before releasing this Japanese version, APTO contacted the original authors directly and received explicit permission to publish this translated version.
We sincerely thank Tinghao Xie and the Sorry-Bench authors for their support and for advancing AI safety research.
Original Dataset Credit
Original Dataset
Sorry-Bench
Original Repository
https://huggingface.co/datasets/sorry-bench/sorry-bench-202503
Citation
If you use this dataset, please cite the original Sorry-Bench dataset.
If you use this Japanese translation, we would also appreciate citation of this repository.
@misc{apto_sorrybench_ja_2026,
title = {APTO-SorryBench-JA},
author = {APTO, Inc.},
year = {2026},
howpublished = {\url{https://huggingface.co/datasets/APTO-001/APTO-SorryBench-JA}},
note = {Japanese translation of Sorry-Bench with permission from the original authors}
}
🇯🇵 日本語
概要
APTO-SorryBench-JAは、AI安全性評価ベンチマークであるSorry-Benchを日本語化・アノテーションしたデータセットです。
Sorry-Benchは、大規模言語モデル(LLM)が有害な依頼を適切に拒否し、安全な依頼には有用な回答を返せるかを評価するためのベンチマークです。
本データセットでは、元の英語プロンプトを保持したまま、日本語訳・日本語参照回答・拒否理由・追加アノテーションを付与しています。
作成プロセス
翻訳・ローカライズ
APTO, Inc.により、元のSorry-Benchを自然な日本語へ翻訳しました。
翻訳にあたっては、評価意図を維持しながら、日本語話者にとって自然な表現となるようローカライズしています。
変更点
元データセットに対し、以下を追加しました。
- 日本語翻訳
- 日本語参照回答
- 拒否理由
- 追加アノテーション(
risk_type、intent、abuse_type、refusal_required)
元の英語プロンプトおよびカテゴリは保持しています。
参照回答
日本語参照回答は、人手で作成・レビューを行いました。
回答は以下を満たすことを目的としています。
- 有害な依頼を適切に拒否する
- 過剰拒否を避ける
- 必要に応じて安全な代替案を提示する
拒否理由
各サンプルには、その依頼を拒否すべき理由(または回答すべき理由)を日本語で記載しています。
LLM-as-a-Judgeや説明可能なAI安全性研究での利用を想定しています。
追加アノテーション
APTO独自に以下のアノテーションを追加しています。
risk_typeintentabuse_typerefusal_required
これらはAPTOのアノテーションガイドラインに基づき、人手で付与・レビューを行いました。
想定用途
- AI安全性研究
- ベンチマーク評価
- モデル評価
- LLM-as-a-Judge
- Safety Fine-tuning
- 日本語AI安全性評価
- 商用・社内評価
本データセットを、有害・違法・非倫理的な目的で利用することは禁止します。
ライセンス
本データセットは、元のSorry-Benchを基に作成しています。
日本語翻訳および追加アノテーションはAPTO, Inc.が作成しました。
公開にあたっては、原著者へ直接連絡し、日本語版公開について明示的な許可をいただいています。
利用にあたっては、元のSorry-Benchライセンスおよび利用条件に従ってください。
謝辞
本データセットは、Sorry-Benchをもとに日本語化したデータセットです。
公開にあたっては、原著者へ直接ご連絡し、日本語版公開について快く許可をいただきました。
Tinghao Xie氏ならびにSorry-Benchの著者の皆様に、心より感謝申し上げます。
元データセット
Original Dataset
Sorry-Bench
Repository
https://huggingface.co/datasets/sorry-bench/sorry-bench-202503
引用
本データセットを利用する場合は、元のSorry-Benchに加えて、本リポジトリも引用いただけますと幸いです。
- Downloads last month
- 2