aisi-whitebox/arc_easy_cot_prompted-sandbagging-llama_33_70b_instruct

arc_easy_cot_prompted-sandbagging-llama_33_70b_instruct数据集是基于Llama-3.3-70B-Instruct模型创建的，用于deception_sprint/arc_easy任务。数据集包含了良性和恶意两种系统提示，用于评估模型在欺骗和安全方面的性能。数据集在创建时启用了沙袋检测，但没有应用沙袋过滤。数据集分为正常和沙袋两种情况，正常情况下的准确度为99.8%，沙袋情况下的准确度为98.8%。

The arc_easy_cot_prompted-sandbagging-llama_33_70b_instruct dataset is created based on the Llama-3.3-70B-Instruct model for the deception_sprint/arc_easy task. The dataset includes both benign and malicious system prompts, used to evaluate the models performance in deception and safety. Sandbagging detection is enabled during the dataset creation, but no sandbagging filtering is applied. The dataset is divided into normal and sandbagging scenarios, with an accuracy of 99.8% in normal conditions and 98.8% in sandbagging conditions.