<![CDATA[The ability to automatically distinguish academic from non-academic texts is an important need in higher education environments. This study aims to design and build a web application that implements the IndoBERT model to classify Indonesian texts into academic or non-academic categories with interpretable output accompanied by classification reasoning. The system is built end-to-end using the PyTorch and HuggingFace Transformers stack for modeling, FastAPI as the backend, and vanilla HTML/CSS/JavaScript for the user interface. The dataset was manually compiled with 209 samples (108 academic and 101 non-academic texts). Model fine-tuning was performed on Google Colab using a Tesla T4 GPU with a learning rate of 2×10⁻⁵, batch size of 16, and eval_loss-based early stopping. The system is also equipped with a zero-shot fallback module so the application remains operational without training. On the internal test set (21 samples), the model achieved 100% accuracy with an F1-score of 1.0000 and an inference speed of 404 samples/second. This research is positioned as a proof-of-concept emphasizing the completeness of the transformer-based NLP pipeline and the interpretability of classification results, rather than a claim of state-of-the-art performance. Discussion of limitations and directions for further development are also presented.]]>
Copyrights © 2026
