AlignmentResearch/HarmBench

该数据集包含了三个配置：default、neg和pos。每个配置都包括了clf_label（分类标签，有两个值：良性Benign和有害Harmful）、instructions（指令）、content（内容）、answer_prompt（回答提示）、proxy_clf_label（代理分类标签）、gen_target（生成目标）和proxy_gen_target（代理生成目标）。default和pos配置的训练集各有200个示例，neg配置的训练集和验证集为空。数据集的下载大小和实际大小也进行了说明。

The dataset consists of three configurations: default, neg, and pos. Each configuration includes clf_label (classification label with two values: Benign and Harmful), instructions, content, answer_prompt, proxy_clf_label, gen_target, and proxy_gen_target. The default and pos configurations have 200 examples in the training set each, while the neg configuration has no examples in both the training and validation sets. The download size and actual size of the dataset are also specified.