日本語ページはこちら
News
- AnswerCarefully Dataset Version 2.0 (ACv2) released (2024/9/12)
- AnswerCarefully Dataset Version 1.0 (ACv1) released (2024/4/30)
Dataset Overview
AnswerCarefully is an instruction dataset specifically aimed at ensuring safety and appropriateness of LLM output in Japanese. This dataset consists of 946 (ACv1) and 1,800 (ACv2) pairs of questions and reference (safe) responses based on the extensive safety taxonomy proposed in Do-Not-Answer dataset. Unlike Do-Not-Answer, the questions and answers in AnswerCarefully are manually created by experienced annotators, reflecting Japanese social/cultural factors.
Description
- Our latest version is Version 2.0 (ACv2), consisting of ACv1 data with corrections as well as newly added data. Corrections include:
- Updated safety taxonomy reflecting Japanese data. See here for details.
- Deleted the questions asking about the system itself that can be answered safely but differently depend on the system, for example “who crated you?” from the risk area of Human-Chatbot Interaction Farms
- Corrected the mistakes in category assignment and misspellings
- AnswerCarefully dataset consists of all original questions and reference (safe) responses
- We inherited the data taxonomy from Do-Not-Answer, but the actual data samples are created from scratch by our seasoned annotators – this is because the data samples in Do-Not-Answer include many templatic, abstract questions; they also do not cover the safety-related topics thar are particularly relevant in Japan.
- Questions include the topics that are of particular significance in Japan (e.g., specific types of crimes, common biases, or dangerous activities)
- Reference (safe) responses are sample answers that are not only safe and appropriate but also helpful when possible.
- We inherited the data taxonomy from Do-Not-Answer, but the actual data samples are created from scratch by our seasoned annotators – this is because the data samples in Do-Not-Answer include many templatic, abstract questions; they also do not cover the safety-related topics thar are particularly relevant in Japan.
- Data Taxonomy follows that of Do-Not-Answer dataset, consisting of 5 risk areas, 12 harm types and 61 subcategories. ACv2 modifies it slightly to accommodate our dataset specific in Japan, resulting in 56 subcategories instead of 61. See here for the detail of the taxonomy update in ACv2.
- ACv2 includes 1,800 question-response pairs with at least 20 samples per each subcategory.
- Of the 1,800 samples, 336 are labeled as test (6 samples from each subcategory) and the remaining 1,464 are labeled as dev.
- Section 6.1 of our paper describes the background of the development of this dataset as well as the experimental results on using the dataset (ACv1) for the alignment of a Japanese LLM via fine-tuning.
Using AnswerCarefully dataset
AC dataset is created solely for the purpose of improving the safety and appropriateness of LLM output in Japanese and other languages, and can be used by anyone for that purpose.
How to download AnswerCarefully dataset
- You will receive an email with the link to the dataset by filling out this form. Any other way of accessing the AC dataset is prohibited. Do not re-distribute the dataset without each user filling in the form above.
- It is prohibited to distribute any derivative data created from AC dataset, including translated data or any synthetic data derived using this dataset.
- This dataset contains offensive, unsafe or inappropriate expressions. Use the dataset with that in mind, and use it only for improving the LLM safety.
- The dataset is in Japanese, including all meta-data (category tags).
Acknowledgements
The development of AnswerCarefully dataset is carried out at LLM-jp, a consortium of research and industry organizations aiming to develop an open-source state-of-the-art Japanese LLM, hosted by the National Institute of Informatics (NII). ACv2 was created at the Center for Research and Development of Large Language Models (LLMC), while ACv1 at Riken AIP (with help from Citadel AI).
Contact：ac_dataset (at) nii.ac.jp
Disclaimers
The creator of this data shall not be responsible for any damage to the user or a third party. In addition, the creator shall not be responsible for any damage to the users or third parties due to delays, interruptions, or suspensions of the provision of this data service. The creator may suspend or discontinue the service of this data or modify the information contained in this data without prior notice.
本データの制作者は、利用者が利用者自身又は第三者に与えた損害について、一切の責任を負わないものとする。また、本データのサービス提供の遅延、中断又は停止により利用者又は第三者が被った損害について、制作者は一切の責任を負わないものとする。制作者は、予告なしに、本データの運営を停止若しくは中止し、又は本データに掲載される情報の全部若しくは一部を変更する場合がある。