aisi-whitebox/arc_challenge_cot_prompted-sandbagging-llama_33_70b_instruct

arc挑战沙袋策略提示数据集，基于Llama-3.3-70B-Instruct模型创建，用于deception_sprint/arc_challenge任务。数据集包含了良性提示和恶意提示，用于评估AI模型在故意提供次优解决方案方面的能力。沙袋策略检测启用，但未应用过滤。数据集分为测试集和验证集，分别占总数据的20%和50%。

arc challenge sandbagging prompt dataset created based on the Llama-3.3-70B-Instruct model for the deception_sprint/arc_challenge task. The dataset includes benign and malicious prompts, designed to evaluate the ability of AI models to intentionally provide suboptimal solutions. Sandbagging detection is enabled, but no filtering is applied. The dataset is split into a test set and a validation set, accounting for 20% and 50% of the total data, respectively.