BullshitEval
收藏资源简介:
BullshitEval数据集是一项专门为评估大型语言模型(LLM)中“机器废话”现象而设计的新基准。该数据集由2400个场景组成,跨越100个不同的AI助手角色,涵盖了广泛的咨询任务和实际应用。每个场景都明确定义了助手的角色,通常涉及产品或服务推荐,或提供专家指导。BullshitEval旨在帮助研究人员更好地理解LLMs中不真实行为的具体表现形式,并评估不同的训练和推理策略对这些行为的影响。
BullshitEval is a novel benchmark specifically developed to evaluate the phenomenon of "machine-generated bullshit" in Large Language Models (LLMs). This dataset includes 2400 scenarios spanning 100 distinct AI assistant roles, covering a wide range of consulting tasks and real-world applications. Each scenario clearly defines the assistant's role, which typically entails product or service recommendations, or the delivery of expert guidance. BullshitEval aims to help researchers gain a better understanding of the specific manifestations of untruthful behaviors in LLMs, and evaluate the impact of various training and inference strategies on such behaviors.




