Gemma Scope Sparse Autoencoder Features

该数据集包含了167个欺骗性特征和269个非欺骗性特征，用于测试人工智能代理在生成欺骗性解释时的可扩展性。数据集包含了每个特征高激活和最小激活的示例，这为详细分析人工智能代理的欺骗行为提供了可能。总体而言，这是一个大型的数据集，总共包含436个特征（167个欺骗性和269个非欺骗性特征），其任务是对人工智能代理在生成欺骗性解释的同时规避检测的能力进行评估。

This dataset contains 167 deceptive features and 269 non-deceptive features, which is designed to test the scalability of AI Agents when generating deceptive explanations. The dataset includes examples of both high activation and minimal activation for each feature, which enables in-depth analysis of the deceptive behaviors of AI Agents. Overall, this is a large-scale dataset with a total of 436 features (167 deceptive and 269 non-deceptive features), whose purpose is to evaluate the ability of AI Agents to generate deceptive explanations while evading detection.