ToMATO
收藏资源简介:
ToMATO是由NTT公司开发的一个心理理论(ToM)基准数据集,旨在通过LLM-LLM对话生成多样化的心理状态数据。该数据集包含5400个问题、753个对话和15种人格特质模式,涵盖了信念、意图、欲望、情感和知识五类心理状态。数据集的生成过程通过信息不对称的设计,促使角色扮演的LLM在对话中表达其心理状态,从而生成虚假信念。ToMATO的应用领域主要集中在评估大型语言模型的心理理论能力,尤其是对虚假信念的理解和对多样化人格特质的鲁棒性。该数据集为研究LLM在真实社交场景中的表现提供了重要参考。
ToMATO is a theory of mind (ToM) benchmark dataset developed by NTT Corporation, which aims to generate diverse mental state data via LLM-to-LLM conversations. This dataset contains 5,400 questions, 753 dialogues, and 15 personality trait patterns, covering five categories of mental states: beliefs, intentions, desires, emotions, and knowledge. The dataset is generated through an information asymmetry design, which prompts role-playing LLMs to express their mental states during conversations, thereby generating false beliefs. The main application areas of ToMATO focus on evaluating the theory of mind capabilities of large language models, particularly their understanding of false beliefs and robustness to diverse personality traits. This dataset provides an important reference for researching the performance of LLMs in real social scenarios.




