PBChat: Enhance Student's Problem Behavior Diagnosis with Large Language Model

Student's problem behaviors are undesirable behaviors encompass actions that deviate from established school standards, potentially impacting students' overall well-being and academic success significantly. Diagnosing these behaviors demands a multidisciplinary understanding, posing a challenge for conventional educators. Capitalizing on the advancements in Large Language Model (LLM) technology, we introduce this PBChat model, a specialized LLM designed for pinpointing problem behaviors. We articulate a theoretical framework for problem behavior diagnosis, laying the conceptual groundwork for PBChat. To train PBChat, we curate a multi-turn dialogue dataset based on annotated cases, and subsequently, fine-tune the ChatGLM2 base model using the QLoRA algorithm to build PBChat model. Experimental assessments gauge the performance of PBChat, with both automated and human evaluations revealing its efficacy in successfully diagnosing problem behaviors, surpassing the capabilities of general LLMs.