【我遇到的问题】 • 现象：该数据集的下载链接已失效 【相关信息】 • 可考虑访问这个链接获取类似文件~https://www.selectdataset.com/dataset/3688356173feccbcf1f1e490ddc6bc72
Catastrophic Response Prompts
收藏arXiv2025-09-30 收录
下载链接：
https://github.com/joaopfonseca/outputscouting
下载链接
链接失效反馈官方服务：
资源简介：
该数据集旨在测试大型语言模型（LLM）在面对可能导致重大损害的简单是非问题时是否会出现灾难性失败。它包含了一些来自现有基准的修改后提示，突显了由于可能出现的灾难性回应不断发展，需要对模型进行持续审计的重要性。该数据集规模较小，仅包含测试提示，其任务是审计LLM在面对灾难性回应时的表现。
This dataset is designed to test whether Large Language Models (LLMs) will suffer catastrophic failures when confronted with simple yes-no questions that could result in severe harm. It includes modified prompts sourced from existing benchmarks, highlighting the growing need for continuous auditing of models amid the evolving risk of catastrophic model outputs. This dataset is small in scale and only contains test prompts, whose core task is to audit the performance of LLMs when they produce catastrophic responses.
