Harmful Behavior Problems (HBP)

该数据集包含了520个恶意问题，这些问题是商业大型语言模型禁止用户查询的，其中包括请求协助犯罪活动或网络攻击的内容。该数据集用于评估IntentObfuscator方法的有效性，并包含了各种有害指令。其规模为520个恶意查询，任务是对语言模型上的越狱攻击方法进行验证。

This dataset includes 520 malicious queries that are prohibited from being submitted by users on commercial large language models (LLMs), covering requests for assistance in criminal activities or cyberattacks. It is employed to evaluate the effectiveness of the IntentObfuscator method and encompasses various harmful instructions. This dataset, which has a total of 520 malicious queries, is designed to validate jailbreaking attack methods on language models.

搜集汇总

背景与挑战

背景概述

该数据集名为'Harmful Behavior Problems (HBP)'，是一个包含521个有害行为请求和响应的CSV文件，用于评估语言模型在面对安全威胁时的表现。其特点包括覆盖网络安全、物理危害、社会危害和金融犯罪等多个有害领域，每个条目由'goal'（有害请求）和'target'（模型响应）组成，结构清晰，适用于对抗性基准测试或AI安全研究。