imerad-kv/r_judge_labelled
收藏资源简介:
该数据集通过LLM法官自动生成的安全标签对R-Judge基准进行了增强。R-Judge是一个用于评估多轮代理场景中LLMs安全判断能力的基准,涵盖五个应用领域(应用、金融、物联网、程序、网络)。数据集包含568行数据(去重后),并提供了基础数据和带有LLM法官标签的增强数据。基础数据列包括应用领域、文件路径、场景ID、场景名称、用户目标、系统提示、对话内容、安全标签(0=安全,1=不安全)、风险描述和攻击类型。增强数据列包括LLM法官的安全标签、置信度、原始输出、解析错误和完整提示。
This dataset augments the R-Judge benchmark with automated safety labels produced by an LLM judge. R-Judge is a benchmark for evaluating the safety judgment capability of LLMs in multi-turn agent scenarios, spanning five application domains (Application, Finance, IoT, Program, Web). The dataset contains 568 rows (after deduplication) and provides both the base data and the augmented data with LLM-judge labels. The base columns include application domain, file path, scenario ID, scenario name, user goal, system prompt, conversation contents, safety label (0=safe, 1=unsafe), risk description, and attack type. The augmented columns include LLM-judge safety label, confidence score, raw output, parse error, and full prompt.




