A Multi-Modal Dataset for Teacher Behavior Analysis in Offline Classrooms

Teacher behavior analysis is essential for enhancing teaching quality and advancing educational development. However, publicly available datasets specifically focused on teacher behavior are scarce, hindering research in this domain. Existing datasets often rely on open course videos from the Internet, which lack the complexity and authenticity of real-world classroom environments and fail to capture teachers' behavioral patterns accurately. Here, we present MM-TBA, a comprehensive multi-modal dataset designed for analyzing teacher behavior in offline classroom settings. Specifically, we recorded 4,839 teaching videos and manually filtered approximately 32,000 seconds of footage, encompassing the instructional activities of over 300 trainee teachers. Based on these videos, we developed a teaching action detection sub-dataset for detecting teachers' temporal actions and an evaluation report sub-dataset for teacher lectures. Additionally, we constructed an instructional design sub-dataset. MM-TBA aims to fill existing gaps and promote scientific research on teacher behavior and cognitive science. We hope that MM-TBA will provide a new research tool for educational science, enabling interdisciplinary applications by combining artificial intelligence with educational technology.