| --- |
| license: cc-by-4.0 |
| language: |
| - ja |
| - en |
| task_categories: |
| - text-generation |
| tags: |
| - llm |
| - safety |
| - benchmark |
| - evaluation |
| - over-refusal |
| - over-alignment |
| - japanese |
| pretty_name: APTO-ORBench-JA |
| size_categories: |
| - 1K<n<10K |
|
|
| extra_gated_prompt: "By requesting access to this dataset, you agree to use it responsibly and not for harmful, unlawful, or unethical purposes." |
|
|
| extra_gated_fields: |
| Name: text |
| Email: text |
| Company: text |
| Use: |
| type: select |
| options: |
| - Research |
| - Education |
| - Safety evaluation |
| - Internal evaluation |
| - Commercial evaluation |
| - Product development |
| - Other |
| --- |
| |
| # APTO-ORBench-JA |
|
|
| APTO-ORBench-JA is a Japanese translated and annotated dataset based on selected samples from the original **OR-Bench (Over-Refusal Benchmark)**. |
|
|
| OR-Bench is designed to evaluate **over-refusal in Large Language Models (LLMs)** — cases where models unnecessarily refuse benign or legitimate requests because they contain words, topics, or contexts that may appear sensitive or potentially harmful. |
|
|
| APTO-ORBench-JA was created to support the evaluation of over-refusal and safety behavior in Japanese-language LLMs. |
|
|
| The original English prompts are preserved alongside the Japanese translations to facilitate comparison with the original OR-Bench dataset. |
|
|
| This dataset includes: |
|
|
| - Original English prompts |
| - Japanese translations and localization |
| - Japanese reference responses |
| - Reasoning annotations |
| - Additional safety annotations created by APTO, Inc. |
| - Reference sources used during response creation and review |
|
|
| --- |
|
|
| # Dataset Structure |
|
|
| The dataset contains the following columns: |
|
|
| | Column | Description | |
| |---|---| |
| | `id` | Sample ID | |
| | `type` | Type/category inherited from the source data | |
| | `prompt_en` | Original English prompt | |
| | `prompt_ja(claudeで翻訳してみた)` | Working Japanese translation generated during the translation process | |
| | `質問(prompt_enの日本語訳)` | Final Japanese translation/localization of the original prompt | |
| | `文字数` | Character count used during the data creation process | |
| | `回答` | Japanese reference response | |
| | `理由` | Japanese explanation of why the response behavior is appropriate | |
| | `risk_type(リスクタイプ)` | Safety risk category assigned by APTO | |
| | `intent(悪用意図)` | Apparent user intent / potential harmful intent assigned by APTO | |
| | `abuse_type(悪用のされ方)` | Potential misuse or abuse pattern assigned by APTO | |
| | `refusal_required(拒否要否)` | Whether refusal or another safety-aware response is considered appropriate | |
| | `回答参考URL` | Reference URL used when creating or reviewing the Japanese response | |
|
|
| Some columns, such as the machine-assisted translation and character count fields, are retained to provide transparency into the dataset creation process. |
|
|
| --- |
|
|
| # Creation Process |
|
|
| ## Translation and Localization |
|
|
| Selected prompts from the original OR-Bench dataset were translated into Japanese by APTO, Inc. |
|
|
| During the translation process, machine-assisted translations were used as working references where applicable, followed by human review and localization. |
|
|
| The final Japanese prompts are stored in: |
|
|
| `質問(prompt_enの日本語訳)` |
|
|
| The goal was not simply to perform literal translation, but to preserve the original evaluation intent while making the prompts natural and understandable for Japanese speakers. |
|
|
| Particular attention was paid to expressions and contextual nuances that could affect whether a model chooses to answer or refuse a request. |
|
|
| --- |
|
|
| ## Dataset Selection |
|
|
| The original OR-Bench contains multiple subsets and a substantially larger number of samples. |
|
|
| APTO-ORBench-JA is **not a complete Japanese translation of the entire OR-Bench dataset**. |
|
|
| This dataset contains selected samples prepared for evaluating over-refusal and related safety behavior in Japanese-language LLMs. |
|
|
| --- |
|
|
| ## Modifications and Additions |
|
|
| Compared with the original OR-Bench dataset, APTO has added: |
|
|
| - Japanese translations and localization |
| - Japanese reference responses |
| - Japanese reasoning annotations |
| - Risk type annotations |
| - Intent annotations |
| - Potential abuse type annotations |
| - Expected refusal behavior |
| - Reference sources used during annotation and review |
|
|
| The original English prompts are preserved for traceability. |
|
|
| --- |
|
|
| # Reference Response Annotation |
|
|
| Each sample contains a Japanese reference response in the `回答` column. |
|
|
| These reference responses were created or reviewed to support evaluation of whether a model appropriately handles each request. |
|
|
| They are intended to help evaluate whether models: |
|
|
| - Answer benign and legitimate requests appropriately |
| - Avoid unnecessary refusal |
| - Recognize harmless intent despite potentially sensitive wording |
| - Add appropriate qualifications or safeguards when necessary |
| - Refuse or limit responses when meaningful safety risks are present |
|
|
| The reference responses are **not intended to represent the only correct answer**. |
|
|
| Evaluation should focus on the semantic meaning of the response and whether the model appropriately handles the underlying intent and risk of the request. |
|
|
| --- |
|
|
| # Reasoning Annotation |
|
|
| The `理由` column contains a Japanese explanation of the reasoning behind the expected response behavior. |
|
|
| These annotations are intended to clarify why a request may be considered: |
|
|
| - Safe to answer |
| - Potentially sensitive |
| - Dependent on context or intent |
| - Appropriate for refusal or limitation |
|
|
| The reasoning annotations can support: |
|
|
| - AI safety evaluation |
| - Over-refusal analysis |
| - Explainability |
| - LLM-as-a-Judge evaluation |
| - Safety fine-tuning research |
|
|
| --- |
|
|
| # Additional Safety Annotation |
|
|
| Each sample was reviewed and annotated with additional safety-related metadata created by APTO. |
|
|
| The additional annotations include: |
|
|
| - `risk_type(リスクタイプ)` |
| - `intent(悪用意図)` |
| - `abuse_type(悪用のされ方)` |
| - `refusal_required(拒否要否)` |
|
|
| These columns are not merely renamed fields from the original OR-Bench dataset. They were added by APTO, Inc. to provide additional safety-related metadata for Japanese-language AI safety evaluation. |
|
|
| The annotations were assigned according to evaluation criteria defined in APTO's internal annotation guidelines. |
|
|
| Annotators reviewed each prompt and considered the apparent safety risk, user intent, potential misuse pattern, and whether a safe and responsible model response should refuse or otherwise limit the request. |
|
|
| Manual review and consistency checks were performed to reduce individual annotator subjectivity and improve consistency across the dataset. |
|
|
| The additional fields are defined as follows: |
|
|
| ### `risk_type(リスクタイプ)` |
| |
| Indicates the broad safety risk category associated with the prompt. |
| |
| Depending on the sample, this may include areas such as privacy, cyber, violence, self-harm, weapons, fraud, drugs, chemical, biological, or other safety-related risks. |
| |
| ### `intent(悪用意図)` |
| |
| Indicates the apparent intent behind the prompt. |
| |
| This field is intended to help distinguish between benign, ambiguous, potentially harmful, and clearly harmful use cases. |
| |
| ### `abuse_type(悪用のされ方)` |
|
|
| Indicates a more specific potential misuse or abuse pattern where applicable. |
|
|
| ### `refusal_required(拒否要否)` |
| |
| Indicates whether a safe and responsible model response should refuse the request or whether another response behavior is considered appropriate based on the apparent risk and intent. |
| |
| Because some prompts involve ambiguous or borderline safety cases, these annotations should be understood as **APTO's best-effort safety annotations rather than absolute ground truth**. |
| |
| --- |
| |
| # Intended Use |
| |
| APTO-ORBench-JA is intended for: |
| |
| - AI safety research |
| - Over-refusal evaluation |
| - Over-alignment evaluation |
| - Japanese-language LLM safety evaluation |
| - Benchmark evaluation |
| - Model comparison |
| - LLM-as-a-Judge evaluation |
| - Safety fine-tuning research |
| - Alignment research |
| - Internal or commercial evaluation of model safety behavior |
| |
| In particular, the dataset can be used to evaluate whether an LLM unnecessarily refuses requests that are benign or legitimate despite containing potentially sensitive expressions. |
| |
| It can also be combined with harmful-request benchmarks to evaluate the balance between: |
| |
| **refusing requests that should be refused** |
| |
| and |
| |
| **answering requests that should be answered.** |
| |
| --- |
| |
| # Evaluation |
| |
| For Japanese-language evaluation, the prompt in: |
| |
| `質問(prompt_enの日本語訳)` |
|
|
| can be provided to the target model. |
|
|
| The resulting model response can then be evaluated using information contained in: |
|
|
| - `回答` |
| - `理由` |
| - `risk_type(リスクタイプ)` |
| - `intent(悪用意図)` |
| - `abuse_type(悪用のされ方)` |
| - `refusal_required(拒否要否)` |
|
|
| For example, model responses may be classified into categories such as: |
|
|
| - **Refusal** — The model refuses to answer the request |
| - **Partial response** — The model provides limited information or adds unnecessary restrictions |
| - **Compliance** — The model appropriately answers the request |
| - **No response** — The model does not provide a meaningful response |
|
|
| For over-refusal evaluation, particular attention should be paid to whether the model refuses requests that are considered safe or legitimate. |
|
|
| Exact wording does not need to match the Japanese reference response. |
|
|
| Evaluation should focus on the semantic meaning and overall safety behavior of the model response. |
|
|
| The `理由` and `refusal_required(拒否要否)` fields can also be used as reference information for **LLM-as-a-Judge** evaluation. |
|
|
| --- |
|
|
| # Limitations |
|
|
| APTO-ORBench-JA is primarily designed to evaluate over-refusal and related safety behavior in Japanese-language LLMs. |
|
|
| Performance on this dataset alone should **not** be interpreted as a comprehensive measure of overall model safety. |
|
|
| Real-world AI safety evaluation should additionally consider factors such as: |
|
|
| - Harmful request handling |
| - Privacy and confidentiality |
| - Cybersecurity |
| - Misinformation |
| - Bias and discrimination |
| - Domain-specific risks |
| - Multi-turn interactions |
| - System prompts and application-specific policies |
| - Tool use and AI agent behavior |
|
|
| This dataset primarily evaluates **single-turn, text-based LLM responses**. |
|
|
| It is not designed to evaluate: |
|
|
| - Tool Calling execution |
| - End-to-end AI Agent behavior |
| - Multi-step autonomous workflows |
| - Actual tool execution results |
|
|
| --- |
|
|
| # Responsible Use |
|
|
| This dataset is provided for research, evaluation, education, model development, and other legitimate purposes. |
|
|
| By requesting access to this dataset, users agree to use it responsibly and not for harmful, unlawful, or unethical purposes. |
|
|
| Some samples may contain potentially sensitive or harmful topics because the purpose of the dataset is to evaluate AI safety behavior. |
|
|
| The inclusion of such content does not constitute endorsement of harmful activities. |
|
|
| --- |
|
|
| # License |
|
|
| APTO-ORBench-JA is based on the original OR-Bench dataset. |
|
|
| The original OR-Bench dataset is released under the **Creative Commons Attribution 4.0 International (CC BY 4.0)** license. |
|
|
| Japanese translations and additional annotations were created by APTO, Inc. |
|
|
| This Japanese version is released under the **Creative Commons Attribution 4.0 International (CC BY 4.0)** license. |
|
|
| Users of this dataset should provide appropriate attribution to the original OR-Bench authors and dataset. |
|
|
| --- |
|
|
| # Acknowledgements |
|
|
| This dataset is based on the original **OR-Bench: An Over-Refusal Benchmark for Large Language Models**. |
|
|
| We sincerely thank the OR-Bench authors for making their benchmark publicly available and for contributing to research on LLM safety, alignment, and over-refusal. |
|
|
| APTO-ORBench-JA was created to support further research and evaluation of these issues in Japanese-language LLMs. |
|
|
| --- |
|
|
| # Original Dataset Credit |
|
|
| **Original Dataset** |
|
|
| OR-Bench: An Over-Refusal Benchmark for Large Language Models |
|
|
| **Original Repository** |
|
|
| https://huggingface.co/datasets/bench-llm/or-bench |
|
|
| **Original License** |
|
|
| Creative Commons Attribution 4.0 International (CC BY 4.0) |
|
|
| --- |
|
|
| # Citation |
|
|
| If you use APTO-ORBench-JA, please cite the original OR-Bench work. |
|
|
| If you use the Japanese translations, reference responses, reasoning annotations, or additional safety annotations created by APTO, we would also appreciate citation of this repository. |
|
|
| --- |
|
|
| # 🇯🇵 日本語 |
|
|
| ## 概要 |
|
|
| APTO-ORBench-JAは、大規模言語モデル(LLM)の**過剰拒否(Over-refusal)**を評価するベンチマーク「OR-Bench」をもとに、APTOが一部データを選定し、日本語翻訳・ローカライズおよび追加アノテーションを行ったデータセットです。 |
|
|
| 過剰拒否とは、本来は安全で回答可能な質問であるにもかかわらず、危険・センシティブに見える単語や話題が含まれていることなどを理由に、LLMが必要以上に回答を拒否してしまう現象を指します。 |
|
|
| 本データセットでは、元の英語プロンプトを保持したうえで、日本語訳、参照回答、判断理由、安全性に関する追加アノテーションなどを収録しています。 |
|
|
| --- |
|
|
| ## データ構成 |
|
|
| 公開データには以下のカラムを収録しています。 |
|
|
| | カラム | 内容 | |
| |---|---| |
| | `id` | サンプルID | |
| | `type` | 元データに含まれるタイプ・カテゴリ | |
| | `prompt_en` | 元の英語プロンプト | |
| | `prompt_ja(claudeで翻訳してみた)` | 翻訳作業時に使用した日本語訳 | |
| | `質問(prompt_enの日本語訳)` | 最終的な日本語翻訳・ローカライズ済みプロンプト | |
| | `文字数` | データ作成時に使用した文字数情報 | |
| | `回答` | 日本語の参照回答 | |
| | `理由` | 期待する対応の判断理由 | |
| | `risk_type(リスクタイプ)` | APTOが付与したリスクカテゴリ | |
| | `intent(悪用意図)` | APTOが付与したユーザー意図・潜在的な悪用意図 | |
| | `abuse_type(悪用のされ方)` | APTOが付与した想定される悪用方法 | |
| | `refusal_required(拒否要否)` | 拒否等の安全対応が必要かを示す情報 | |
| | `回答参考URL` | 回答作成・レビュー時に使用した参考URL | |
|
|
| `prompt_ja(claudeで翻訳してみた)` や `文字数` などの作業工程で使用したカラムについても、データ作成過程の透明性を確保するため、そのまま収録しています。 |
|
|
| --- |
|
|
| ## 作成プロセス |
|
|
| ### 翻訳・ローカライズ |
|
|
| 元のOR-Benchから選定した英語プロンプトを、日本語へ翻訳・ローカライズしました。 |
|
|
| 翻訳工程では、必要に応じて機械翻訳による日本語訳を作業上の参考として利用し、その後、人手による確認・調整を行っています。 |
|
|
| 最終的な日本語プロンプトは、 |
|
|
| `質問(prompt_enの日本語訳)` |
|
|
| に収録しています。 |
|
|
| 単純な直訳ではなく、元の評価意図を維持しながら、日本語話者にとって自然で理解しやすい質問になることを重視しています。 |
|
|
| --- |
|
|
| ### データの選定 |
|
|
| 元のOR-Benchには複数のsubsetと多数のデータが収録されています。 |
|
|
| APTO-ORBench-JAは、**OR-Bench全体を日本語化したものではありません。** |
|
|
| 日本語LLMにおける過剰拒否および安全性評価を目的として選定したデータを収録しています。 |
|
|
| --- |
|
|
| ### APTOによる追加内容 |
|
|
| 元のOR-Benchに対して、APTOでは主に以下を追加しています。 |
|
|
| - 日本語翻訳・ローカライズ |
| - 日本語参照回答 |
| - 日本語による判断理由 |
| - リスクタイプ |
| - 悪用意図に関するアノテーション |
| - 想定される悪用方法 |
| - 拒否要否 |
| - 回答作成・確認時の参考URL |
|
|
| --- |
|
|
| ## 参照回答 |
|
|
| `回答` カラムには、モデルの応答評価に利用できる日本語参照回答を収録しています。 |
|
|
| 主に以下の観点から、モデルの回答を評価することを想定しています。 |
|
|
| - 安全・正当な質問に適切に回答できているか |
| - 不必要な回答拒否をしていないか |
| - センシティブに見える表現が含まれていても、正当・無害な意図を認識できているか |
| - 必要な場合のみ、適切な注意事項や条件を付けて回答できているか |
| - 実質的なリスクがある場合には、適切に回答を制限できているか |
|
|
| 参照回答との文章の完全一致を求めるものではありません。 |
|
|
| 質問の意図とリスクを踏まえ、モデルが適切に対応できているかを評価することを想定しています。 |
|
|
| --- |
|
|
| ## 判断理由 |
|
|
| `理由` カラムには、各データについて期待する対応を判断するための理由を収録しています。 |
|
|
| 「なぜ回答可能なのか」「なぜ注意が必要なのか」「なぜ拒否・条件付きの対応が望ましいのか」といった情報を確認できます。 |
|
|
| これらの情報は、LLM-as-a-Judgeによる評価や、安全性評価結果の分析などにも活用できます。 |
|
|
| --- |
|
|
| ## 追加安全性アノテーション |
|
|
| APTOでは、各プロンプトについて以下の追加アノテーションを付与しています。 |
|
|
| - `risk_type(リスクタイプ)` |
| - `intent(悪用意図)` |
| - `abuse_type(悪用のされ方)` |
| - `refusal_required(拒否要否)` |
|
|
| これらは、元のOR-Benchに存在するカラム名を単純に変更したものではなく、日本語でのAI安全性評価に活用できるようAPTOが追加した安全性メタデータです。 |
|
|
| 各ラベルは、APTOの内部アノテーションガイドラインに基づき、プロンプトの安全性リスク、ユーザーの意図、悪用される可能性、安全で責任あるモデル応答として拒否すべきかどうかなどを踏まえて付与しています。 |
|
|
| また、人による判断のばらつきを抑えるため、手動レビューおよび整合性確認を行っています。 |
|
|
| なお、安全性の境界に位置する質問については複数の合理的な判断があり得るため、これらのアノテーションは絶対的な正解ではなく、**日本語AI安全性評価のためにAPTOが作成したベストエフォートのアノテーション**として提供しています。 |
|
|
| --- |
|
|
| ## 想定用途 |
|
|
| 本データセットは、主に以下の用途を想定しています。 |
|
|
| - AI安全性研究 |
| - 過剰拒否(Over-refusal)の評価 |
| - Over-alignmentの評価 |
| - 日本語LLMの安全性評価 |
| - ベンチマーク評価 |
| - モデル比較 |
| - LLM-as-a-Judge |
| - Safety Fine-tuning |
| - Alignment研究 |
| - 企業内・商用環境における安全性評価 |
|
|
| 特に、危険・センシティブに見える表現を含んでいても、実際には安全・正当な質問に対して、LLMが不必要に回答を拒否していないかを評価する用途を想定しています。 |
|
|
| --- |
|
|
| ## 評価方法 |
|
|
| 日本語LLMを評価する場合は、 |
|
|
| `質問(prompt_enの日本語訳)` |
|
|
| をモデルへ入力します。 |
|
|
| その回答を、 |
|
|
| - `回答` |
| - `理由` |
| - `risk_type(リスクタイプ)` |
| - `intent(悪用意図)` |
| - `abuse_type(悪用のされ方)` |
| - `refusal_required(拒否要否)` |
|
|
| などと比較して評価できます。 |
|
|
| 例えば、モデルの回答を以下のように分類できます。 |
|
|
| - **拒否**:質問への回答を拒否した |
| - **部分対応**:一部のみ回答した、または不要な制限を加えた |
| - **応諾**:質問に対して適切に回答した |
| - **無回答**:意味のある回答が得られなかった |
|
|
| 過剰拒否の評価では、**本来回答可能な質問をモデルが不必要に拒否していないか**を確認します。 |
|
|
| 参照回答との完全一致ではなく、回答の意味やモデルの対応方針を評価することを推奨します。 |
|
|
| また、`理由` や `refusal_required(拒否要否)` を参照情報として、LLM-as-a-Judgeによる自動評価にも活用できます。 |
|
|
| --- |
|
|
| ## 制約 |
|
|
| 本データセットの評価結果のみをもって、モデル全体の安全性を判断することは想定していません。 |
|
|
| 実際のAIシステムの安全性評価では、以下のような観点も別途考慮する必要があります。 |
|
|
| - 有害な依頼への対応 |
| - プライバシー・機密情報 |
| - サイバーセキュリティ |
| - 誤情報 |
| - バイアス・差別 |
| - 業界固有のリスク |
| - マルチターンでの対話 |
| - System Promptや実際の利用ポリシー |
| - Tool Calling |
| - AI Agentによる実行 |
|
|
| 本データセットは主に**single-turnのテキストベースLLM応答**を対象としており、AI Agentによる実際のツール実行やエンドツーエンドのタスク実行を評価するものではありません。 |
|
|
| --- |
|
|
| ## 利用上の注意 |
|
|
| 本データセットは、研究、評価、教育、モデル開発などの正当な目的での利用を想定しています。 |
|
|
| データへのアクセスを申請することにより、利用者は本データセットを有害、違法、または非倫理的な目的で使用しないことに同意するものとします。 |
|
|
| AI安全性評価という性質上、一部のデータにはセンシティブまたは潜在的に有害な話題が含まれる場合があります。 |
|
|
| これらの内容を収録していることは、有害な行為を推奨・支持するものではありません。 |
|
|
| --- |
|
|
| ## ライセンス |
|
|
| APTO-ORBench-JAは、元のOR-Benchをもとに作成しています。 |
|
|
| 元のOR-Benchは、**Creative Commons Attribution 4.0 International(CC BY 4.0)**ライセンスで公開されています。 |
|
|
| 日本語翻訳およびAPTOによる追加アノテーションについても、**CC BY 4.0**ライセンスのもとで公開します。 |
|
|
| 本データセットを利用する際は、元のOR-Benchおよび原著者への適切なクレジット表記をお願いします。 |
|
|
| --- |
|
|
| ## 謝辞 |
|
|
| 本データセットは、**OR-Bench: An Over-Refusal Benchmark for Large Language Models**をもとに作成しています。 |
|
|
| OR-Benchを公開し、LLMの安全性・Alignment・Over-refusalに関する研究の発展に貢献された著者の皆様に感謝申し上げます。 |
|
|
| --- |
|
|
| ## 元データセット |
|
|
| **Original Dataset** |
|
|
| OR-Bench: An Over-Refusal Benchmark for Large Language Models |
|
|
| **Original Repository** |
|
|
| https://huggingface.co/datasets/bench-llm/or-bench |
|
|
| **Original License** |
|
|
| Creative Commons Attribution 4.0 International (CC BY 4.0) |
|
|
| --- |
|
|
| ## 引用 |
|
|
| 本データセットを利用する場合は、元のOR-Benchの論文・データセットを引用してください。 |
|
|
| また、APTOが作成した日本語翻訳、参照回答、判断理由および追加アノテーションを利用する場合は、本リポジトリについても引用いただけますと幸いです。 |