longtermrisk/school-of-reward-hacks
收藏官方服务：
资源简介：
School of Reward Hacks数据集包含主要数据集和匹配的控制数据集，用于研究在LLM中无害任务的黑客攻击如何泛化到行为不匹配的情况。数据集中的字段包括用户消息、低质量助手响应、控制响应、任务描述、评估指标和作弊方法。
The School of Reward Hacks dataset includes both the main dataset and a matched control dataset, used to study how hacking harmless tasks generalizes to misaligned behavior in LLMs. The fields in the dataset include user message, low-quality assistant response, control response, task description, evaluation metric, and cheat method.
提供机构：
longtermrisk
