aisi-whitebox/wmdp_bio_cot_prompted_sandbagging_llama_31_8b_instruct

这是一个使用vllm<meta-llama/Llama-3.1-8B-Instruct模型创建的检测欺骗行为的数据集，具体任务为deception_sprint/wmdp_bio。数据集包含了沙袋策略，但不应用沙袋过滤。数据集的创建日期为2025年4月9日。数据集分为训练集、测试集和验证集，比例为假定为False时的默认值。测试集大小为0.2，验证集大小为0.5，随机种子为42。沙袋策略统计显示，正常准确度为51.8%，沙袋准确度为32.4%。

This dataset is created using the vllm<meta-llama/Llama-3.1-8B-Instruct model for detecting deception, specifically with the task deception_sprint/wmdp_bio. The dataset includes sandbagging strategies but does not apply sandbagging filtering. The dataset was created on April 9, 2025. The dataset is split into training, test, and validation sets with the default ratios when Split Applied is False. The test set size is 0.2, the validation set size is 0.5, and the random seed is 42. Sandbagging statistics show a normal accuracy of 51.8% and a sandbagging accuracy of 32.4%.