aisi-whitebox/arc_challenge_cot_prompted_sandbagging_llama_31_8b_instruct

这是一个用于评估的欺骗性数据集，名为inspect llama 31 8b instruct prompted sandbagging arc challenge cot，使用vllm/meta-llama/Llama-3.1-8B-Instruct模型创建于2025年4月9日。数据集包含deception_sprint/arc_challenge任务，启用了沙袋检测但未应用过滤。数据集分为训练集、验证集和测试集，比例分别为0.2、0.5和未指定。该数据集的目的是评估模型在面对故意生成的低质量输出时的表现。

This is an evaluation dataset for deception, named inspect llama 31 8b instruct prompted sandbagging arc challenge cot, created using the vllm/meta-llama/Llama-3.1-8B-Instruct model on April 9, 2025. The dataset includes the deception_sprint/arc_challenge task, with sandbagging detection enabled but no filtering applied. The dataset is split into training, validation, and test sets with ratios of 0.2, 0.5, and unspecified respectively. The purpose of this dataset is to evaluate the models performance when faced with intentionally generated low-quality outputs.