利用規約
- 本データセットは、日本語および他の言語のLLMの安全性を向上させるという目的のため、商用利用も含め公開しています。
- 本データセットを、LLMの安全性対策の回避など、LLMの安全性向上以外の目的で使用することを禁じます。
- 本データの再配布は許可しませんが、派生データ（本データセットと重複しない、翻訳や類似データ、合成プロンプトを含む）の作成や配布を制限するものではありません。ただし派生データの配布時には本データを利用したことを明記してください。
- 本データセットはその性質上有害・不適切な表現を含みます。承諾の上、LLMの安全性向上のためにご使用ください。
免責事項
- 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。

JSocialFact2 Terms of Use

Terms of Use
- This dataset is made publicly available for the purpose of improving the safety of LLMs in Japanese and other languages, including for commercial use.
- Users agree not to use this dataset for any purpose other than improving the safety of LLMs. In particular, it is strictly prohibited to use it to circumvent the safety measures of LLMs.
- Redistribution of this dataset is not allowed. However, we allow the creation and distribution of any derivative data created from it (including translations, similar data, or synthetic prompts) on the condition that (1) the derivative data does not contain the original data in this dataset; (2) an attribution is given to this dataset.
- Due to the nature of this dataset, it contains expressions that may be considered inappropriate, unsafe or offensive. Please use it with caution for the purpose of improving LLM safety.
Disclaimer
- 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.

JSocialFact2

概要

JSocialFact2 は、ソーシャルメディア上で実際に流通した日本語の偽・誤情報を題材にした評価用データセットです。各項目は、その情報を信じた人が LLM に尋ねそうな質問と、それに対する参考回答の組で構成されています。

2025年12月〜2026年1月に収集した X（旧 Twitter）のポストおよびコミュニティノートを元に質問と参考回答を人手で作成し、複数のアノテータによる検証を経て構築しました。各項目には、JSocialFact と同じ、情報の類型・トピック・ AnswerCarefully リスクカテゴリに加えて、新たに、有害性スコアと期待方向性 （事実に照らして回答が取るべき立場）を付与しています。

JSocialFact2 は、JSocialFact（中里ら, 2024）と同じ着想に基づきますが、データは全て新規に収集・作成したものです。

全件を評価用データとして公開しています。学習用の split はありません。

Overview

JSocialFact2 is a Japanese evaluation dataset built from misinformation that actually circulated on social media. Each item pairs a question that a person who believed the misinformation might ask an LLM with a reference answer to that question.

Questions and reference answers were written by hand based on posts and Community Notes on X (formerly Twitter) collected between December 2025 and January 2026, and were verified by multiple annotators. In addition to information type, topic, and AnswerCarefully risk categories, each item carries a harmfulness score and an expected answer direction — the stance an answer shouldtake given the facts.

JSocialFact2 is built on the same idea as JSocialFact (Nakazato et al., 2024) but uses newly collected data and contains no items from ver.1.

All items are released as evaluation data. There is no training split.

Language: Japanese
Intended use: Evaluating the factual accuracy and safety of LLMs. Not for training — training on this data would invalidate it as a benchmark.

Dataset Structure

Split	File	Format	Size
`test`	`data/test.jsonl`	JSONL	190

Fields

Key	Type	Description
`Id`	string	Item ID
`question`	string	A question a person might ask an LLM after encountering the misinformation
`reference-answer`	string	Reference answer written by an annotator
`info-category-01`	string	Type of information (fabrication, conspiracy theory, pseudoscience, etc.)
`topic-01`	string	Topic (daily life, society, science, medicine, etc.)
`harm-type`	string	AnswerCarefully risk category, middle level
`specific-harm`	string	AnswerCarefully risk category, fine level
`harmfulness-score`	int	Harmfulness on a 1–3 scale (see below)
`expected-direction`	string	The stance an answer should take given the facts (see below)
`post-date`	string	Posting date and time of the source X post, in ISO 8601 format with JST offset (e.g., `2026-01-05T10:00:00+09:00`)

`harmfulness-score`

Rated on a three-point scale following the severity levels of PKU-SafeRLHF (Ji et al., 2024), based on the scope of impact and the response measures required.

Value	Level	Impact	Response measures
1	Minor	Slight personal discomfort or limited societal impact	Personal measures are enough
2	Moderate	Noticeable damage to individuals, hard to undo in the short term	Requires government or professional intervention
3	Severe	Targets groups; wide-reaching and long-lasting serious consequences	Requires national or international resources

`expected-direction`

Value	Meaning
`yes`	The premise of the question should be affirmed
`no`	The premise of the question should be denied
`neither`	Conditional or partial; neither affirmation nor denial applies
`unknown`	Scientifically unsettled, or there is not enough evidence to answer

Construction

Collected posts containing misinformation from X posts and Community Notes
For each post, wrote by hand a question a believer might ask and a reference answer
Automatic filtering with an LLM (four criteria for questions, four for answers)
Manual verification by multiple annotators
Re-checked items where annotators disagreed or where labels needed correction

Usage

from datasets import load_dataset

# Load the test split.
test = load_dataset("llm-jp/jsocialfact2", split="test")

License

See the LICENSE file.

How to cite

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

Data Card Authors

The names are listed in alphabetical order.

Ryuhei Miyazato

Downloads last month: 17

Total file size:

167 kB

Paper for llm-jp/JSocialFact2

PKU-SafeRLHF: A Safety Alignment Preference Dataset for Llama Family Models

Paper • 2406.15513 • Published Jun 20, 2024 • 1