<![CDATA[The rapid growth of user-generated content on digital platforms has increased the difficulty of moderating hate-related expressions, particularly in linguistically diverse environments such as Indonesian social media. Automated hate speech detection systems are therefore expected to operate not only with reliable predictive behavior but also with practical efficiency under large-scale deployment conditions. This study reports an applied evaluation of Transformer-based models for multi-label hate speech detection on Indonesian digital platforms. Rather than introducing a new classification architecture, the work focuses on assessing multiple pretrained language models within a unified and reproducible evaluation framework. The analysis examines overall model behavior, per-label performance tendencies, inference efficiency, and common error patterns under realistic multi-label settings. The results indicate that IndoBERT-based models (indobenchmark/indobert-base-p1 and cahya/bert-base-indonesian) achieved the strongest predictive performance for multi-label hate speech detection, although performance differences across the evaluated Transformer models remained relatively incremental. Experimental results show that the best-performing model achieved a macro-F1 score of 0.9742 and a micro-F1 score of 0.9749, while other Transformer models demonstrated competitive performance with macro-F1 values ranging from 0.955 to 0.964. In terms of efficiency, distilled models provided faster inference (approximately 5–7 ms per sample) compared to full-size models (8–11 ms per sample), highlighting a practical trade-off between predictive performance and computational cost. These findings emphasize the importance of practical evaluation strategies and suggest that flexible model configurations are more suitable than reliance on a single high-capacity model in real-world moderation systems.]]>
