jordan-taylor-aisi/test_cot

这是一个用于评估和检测沙袋策略的数据集，基于Llama-3.1-8B-Instruct模型，针对wmdp_bio任务。数据集包含良性提示和指导模型故意提供次优解决方案的恶意提示，用于测试模型在特定任务上的表现和欺骗性。数据集不进行分割，但有专门的测试集和验证集，以及相关的统计信息。

This is a dataset for evaluation and detection of sandbagging strategies based on the Llama-3.1-8B-Instruct model for the wmdp_bio task. The dataset includes both benign prompts and malicious prompts designed to guide the model to deliberately provide suboptimal solutions, testing the models performance on a specific task and its deceptiveness. The dataset is not split but has dedicated test and validation sets, along with relevant statistics.