AnswerCarefully

# AnswerCarefully

概要

AnswerCarefullyは日本語LLM 出力の安全性・適切性に特化したインストラクションデータセットです。
このデータセットは、英語の要注意回答を集めた [Do-Not-Answer データセット](https://github.com/Libr-AI/do-not-answer) の包括的なカテゴリ分類に基づき、人手で質問・回答ともに日本語サンプルを集めたオリジナルのデータセットです。
データセットの詳細については、[こちら](https://llmc.nii.ac.jp/answercarefully-dataset/)をご覧ください。

Overview

AnswerCarefully is an instruction dataset specifically aimed at ensuring safety and appropriateness of LLM output in Japanese.
This dataset consists of original pairs of questions and reference (safe) responses based on the extensive safety taxonomy proposed in [Do-Not-Answer dataset](https://github.com/Libr-AI/do-not-answer).
The questions and answers in AnswerCarefully are manually created by experienced annotators, reflecting Japanese social/cultural factors. For more information, refer to the [dataset homepage](https://llmc.nii.ac.jp/en/answercarefully-dataset/).

## Usage

```python
from datasets import load_dataset

# load dev and test splits of v2.2 (latest version)
v2_2_dev = load_dataset("llm-jp/AnswerCarefully", "v2.2", split="dev")
v2_2_test = load_dataset("llm-jp/AnswerCarefully", "v2.2", split="test")

# load dev and test splits of v2.0
v2_dev = load_dataset("llm-jp/AnswerCarefully", "v2.0", split="dev")
v2_test = load_dataset("llm-jp/AnswerCarefully", "v2.0", split="test")

# load dev and test splits of v1.0
v1_dev = load_dataset("llm-jp/AnswerCarefully", "v1.0", split="dev")
v1_test = load_dataset("llm-jp/AnswerCarefully", "v1.0", split="test")
```

## Send Questions to

ac-dataset(at)nii.ac.jp

## License

See the [LICENSE](LICENSE) file.

## How to cite

If you find our work helpful, please feel free to cite the paper.

```
@misc{suzuki2025answercarefullydatasetimprovingsafety,
title={AnswerCarefully: A Dataset for Improving the Safety of Japanese LLM Output},
author={Hisami Suzuki and Satoru Katsumata and Takashi Kodama and Tetsuro Takahashi and Kouta Nakayama and Satoshi Sekine},
year={2025},
eprint={2506.02372},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2506.02372},
}
```

## Model Card Authors

*The names are listed in alphabetical order.*

Hirokazu Kiyomaru, Takashi Kodama, and Hisami Suzuki.

# AnswerCarefully

## 概要
AnswerCarefully是一款专门针对日语大语言模型（LLM）输出内容的安全性与合规性打造的指令数据集。本数据集基于收录英语高危应答的[Do-Not-Answer数据集（Do-Not-Answer Dataset）](https://github.com/Libr-AI/do-not-answer)所提出的全面安全分类体系，由人工手动收集并构建了日语版的问答样本对，属于原创性数据集。有关数据集的详细信息，请参阅[此页面](https://llmc.nii.ac.jp/answercarefully-dataset/)。

## 概述
AnswerCarefully是一款专为保障日语大语言模型（LLM）输出内容的安全性与合规性而设计的指令数据集。本数据集基于[Do-Not-Answer数据集（Do-Not-Answer Dataset）](https://github.com/Libr-AI/do-not-answer)所提出的全面安全分类体系，构建了原创的问题与参考（安全）应答样本对。AnswerCarefully中的所有问答样本均由经验丰富的标注人员手动创建，充分体现了日本社会与文化特征。如需了解更多详细信息，请参阅[数据集主页](https://llmc.nii.ac.jp/en/answercarefully-dataset/)。

## 使用方法

python
from datasets import load_dataset

# load dev and test splits of v2.2 (latest version)
v2_2_dev = load_dataset("llm-jp/AnswerCarefully", "v2.2", split="dev")
v2_2_test = load_dataset("llm-jp/AnswerCarefully", "v2.2", split="test")

# load dev and test splits of v2.0
v2_dev = load_dataset("llm-jp/AnswerCarefully", "v2.0", split="dev")
v2_test = load_dataset("llm-jp/AnswerCarefully", "v2.0", split="test")

# load dev and test splits of v1.0
v1_dev = load_dataset("llm-jp/AnswerCarefully", "v1.0", split="dev")
v1_test = load_dataset("llm-jp/AnswerCarefully", "v1.0", split="test")

## 咨询方式

ac-dataset(at)nii.ac.jp

## 许可证

详见[LICENSE](LICENSE)文件。

## 引用方式

若您认为本研究对您有所助益，请随时引用该论文。

@misc{suzuki2025answercarefullydatasetimprovingsafety,
title={AnswerCarefully: A Dataset for Improving the Safety of Japanese LLM Output},
author={Hisami Suzuki and Satoru Katsumata and Takashi Kodama and Tetsuro Takahashi and Kouta Nakayama and Satoshi Sekine},
year={2025},
eprint={2506.02372},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2506.02372},
}

## 模型卡片作者

*姓名按字母顺序排列。*

Hirokazu Kiyomaru, Takashi Kodama, and Hisami Suzuki.