Knowledge-Jailbreak Dataset

该数据集是一个大规模的数据集，包含了12,974个知识逃逸对，用于评估大型语言模型（LLM）在运用领域知识时的安全性。数据集分为已知领域和未知领域，其中未知领域包含6个领域的91个数据点，用于测试。已知领域则按照8:2的比例分为训练集和测试集。该数据集的任务是通过知识驱动的逃逸生成来评估LLM的安全性。

This is a large-scale dataset comprising 12,974 knowledge escape pairs, developed to evaluate the safety of Large Language Models (LLMs) when leveraging domain-specific knowledge. The dataset is divided into known domains and unknown domains. The unknown domains include 91 data points across 6 domains for testing LLMs. The known domains are split into training and test sets at an 8:2 ratio. The task of this dataset is to evaluate the safety of LLMs via knowledge-driven escape generation.

搜集汇总

背景与挑战

背景概述

Knowledge-Jailbreak Dataset 是一个用于研究知识驱动越狱攻击的数据集，基于论文'Knowledge-to-Jailbreak: One Knowledge Point Worth One Attack'开发。它包含原始提示、知识提示和突变越狱提示，支持通过代码流程生成训练数据，并提供了测试不同模型越狱有效性的实验工具，旨在利用特定知识点攻击大语言模型的安全性。