HPAI-BSC/Egida

Egida数据集是一个从不同外部来源收集的不安全请求的扩展集合。该数据集通过人工细粒度主题分类和多种越狱攻击方法的运用进行了增强和扩展。数据集总共包含了2949个危险问题或指令，这些问题是从九个不同的公共数据集中汇编而成，并经过人工审查和去重。数据集还包括了通过越狱攻击扩展后的61330个不安全实例。此外，Egida数据集还有三个不同的扩展版本，用于实验和评估LLM模型在越狱攻击存在下的安全性。

The Egida dataset is an expanded collection of unsafe requests gathered from a variety of external sources. The dataset is enhanced and extended through manual fine-grained topic classification and the application of various jailbreaking attacks. It contains a total of 2,949 dangerous questions or instructions compiled from nine different public datasets, which have been manually reviewed and deduplicated. The dataset also includes 61,830 unsafe instances that have been expanded using jailbreaking attacks. In addition, there are three different extended versions of the Egida dataset for experimenting and evaluating the safety of LLMs in the presence of jailbreaking attacks.