The paper introduces COM² (COMplex Commonsense), a new dataset designed to enhance complex commonsense reasoning in language models. COM² is created by sampling multi-hop logical queries from a commonsense knowledge graph (CSKG) and verbalizing them into multiple-choice and text generation questions using handcrafted rules and large language models. The dataset aims to address the challenge of reasoning about complex event interactions, which is difficult for language models due to data scarcity. Experiments show that models trained on COM² significantly improve their complex reasoning abilities, leading to better zero-shot performance in both in-domain and out-of-domain tasks for question answering and generative commonsense reasoning. The paper also discusses the construction process, including pre-processing, query sampling, and verbalization, and presents experimental results demonstrating the effectiveness of COM² in improving model performance.The paper introduces COM² (COMplex Commonsense), a new dataset designed to enhance complex commonsense reasoning in language models. COM² is created by sampling multi-hop logical queries from a commonsense knowledge graph (CSKG) and verbalizing them into multiple-choice and text generation questions using handcrafted rules and large language models. The dataset aims to address the challenge of reasoning about complex event interactions, which is difficult for language models due to data scarcity. Experiments show that models trained on COM² significantly improve their complex reasoning abilities, leading to better zero-shot performance in both in-domain and out-of-domain tasks for question answering and generative commonsense reasoning. The paper also discusses the construction process, including pre-processing, query sampling, and verbalization, and presents experimental results demonstrating the effectiveness of COM² in improving model performance.