With the continuous development of open-source communities, the variety of open-source software projects has also been steadily increasing. However, existing data in these communities, such as tags and descriptions, cannot be used to directly infer the fine-grained domains of projects. As a result, developers and users need to spend unnecessary time manually identifying the categories of open-source software projects, and open-source researchers may be unable to effectively distinguish fine-grained categories in their project samples, thus affecting the specificity of research results. To address these issues, this study proposes an interactive distillation method that can automatically classify open-source projects into 12 categories, including code development tools or plugins, Web applications, and other categories. First, 30310 open-source software projects are annotated in this study. Then, existing large model methods and machine learning methods are tested on the fine-grained domain classification task. Finally, based on the problems identified in the baseline methods, this study distills keyword features from large models and further optimizes the model’s classification capability by introducing a collaboration method between large and small models. Validation results on the annotated dataset show that the proposed method achieves an average Precision of 92.1%, an average Accuracy of 91.8%, and an average F1-measure of 91.8% after ten-fold cross-validation. Compared with the baseline methods, the proposed method improves Precision by 26.3%–32.8%, Accuracy by 24.5%–39.9%, and F1-measure by 26.1%–40.7%.