Large language models are increasingly being adopted in educational settings for their exceptional conversational abilities and vast knowledge base. Nevertheless, when assessing students’ emotional status and learning efficiency through semi-structured counseling evaluations in academic contexts, the responses and assessments generated by large language models often lack specificity and practical relevance. To address this limitation, we propose EduMate, a two-stage model designed to interact and evaluate students’ emotions and learning efficiency through a semi-structured counseling paradigm. For building EduMate, we firstly collaborate with the on-campus academic counselors to construct a semi-structured real-world academic counseling dialogue dataset in Chinese, accompanied by 16 assessment dimensions. EduMate is then fine-tuned on this dataset, and its performance is compared to the current large language models in processing semi-structured academic counseling dialogues, as well as their effectiveness in assessing students’ learning status based on dialogue content and predefined evaluation dimensions. Our findings reveal that EduMate achieves significant improvements in the semi-structured academic counseling dialogues, outperforming GPT-4 by more than 50%. By employing fine-grained retrieval of dialogue history, EduMate further demonstrates an 8% increase in accuracy over GPT-4 in evaluating students’ emotional status and learning efficiency. Our data is freely available for research use.