<![CDATA[Court decision documents in Indonesia are generally lengthy in structure and not yet digitally organized, making the automatic extraction of information a challenging task. Furthermore, the application of deep learning models to legal documents is often constrained by high computational demands. Therefore, this study aims to develop an efficient classification system for legal decision structures (IndoLaw) using the IndoBERT model with the Low-Rank Adaptation (LoRA) method. The study utilizes the IndoLaw dataset consisting of 111,108 rows of first-level decision text. Data splitting was performed using a document-level split with an 80:10:10 ratio to prevent data leakage across documents. The LoRA method was applied to the attention layers of IndoBERT by inserting low-dimensional matrices, making the fine-tuning process more efficient. The model was then integrated into a Streamlit-based application to support interactive text classification and PDF document extraction. Test results show that the IndoBERT + LoRA model achieved an Accuracy and F1-Score of 0.9944, outperforming the Full Fine-Tuning method which recorded an Accuracy of 0.9939 and an F1-Score of 0.9938. In addition, the use of LoRA reduced trainable parameters by up to 99.74%, decreased GPU VRAM usage by 44% (from 2.5 GB to 1.4 GB), and accelerated computation time per epoch by 36%. Based on these results, the IndoBERT + LoRA approach is proven capable of producing accurate and efficient legal document classification with lower computational requirements.]]>
Copyrights © 2026
