cracklinoatbran/reward_hacking_monitor_2046
收藏资源简介:
该数据集名为reward_hacking_monitor_2046,主要用于监测无害任务中的奖励黑客行为。数据集包含(user, response)对以及真实标签,用于测试监控器是否能正确识别助理是否进行了奖励黑客行为。数据集来源于longtermrisk/school-of-reward-hacks,包含1,073个黑客变体和973个控制变体,总计2,046行。每个条目包含多个字段,如eval_row_id、task、evaluation_metric等。数据集用于评估监控器的性能,包括掩盖率、误报率等指标。
The dataset named reward_hacking_monitor_2046 is designed to monitor reward-hacking behavior on harmless tasks. It consists of (user, response) pairs along with ground-truth labels, used to test whether the monitor can correctly identify if the assistant engaged in reward hacking. The dataset is derived from longtermrisk/school-of-reward-hacks, containing 1,073 hack variants and 973 control variants, totaling 2,046 rows. Each entry includes multiple fields such as eval_row_id, task, evaluation_metric, etc. The dataset is used to evaluate the monitors performance, including metrics like cover-up rate and false-positive rate.




