<![CDATA[Social media has become a primary platform for Indonesian society to interact and exchange information online. However, freedom of expression in digital spaces is often misused through the use of harsh, offensive, and hateful language. This study aims to develop a toxic comment classification model for the Indonesian language using the IndoBERT architecture through a fine-tuning process. IndoBERT was selected for its capability to understand bidirectional semantic context and its pretraining on a Bahasa Indonesia corpus, making it suitable for handling informal language styles, abbreviations, and common code-mixing phenomena in social media texts. The dataset used in this study is the Indonesian Abusive and Hate Speech Twitter Text, consisting of 12,942 entries 11,647 for training and 1,295 for validation. The research was conducted online using Google Colaboratory with GPU acceleration. The research stages included data preprocessing, tokenization, model training, and evaluation using precision, recall, F1-score, and confusion matrix as metrics. Evaluation results show that the fine-tuned IndoBERT model achieved high performance, with an average precision of 0.8842, recall of 0.884, F1-score of 0.883, and accuracy of 0.8834. These results indicate balanced performance across classes and strong model stability in detecting both toxic and non-toxic comments. This study contributes to the development of an automated Indonesian-language content moderation system, which can be deployed as a comment detection module via API. Although limited to Twitter data and binary classification, this model has the potential to be extended toward multi-class and cross-platform classification in supporting safer and healthier digital spaces in Indonesia.]]>
Copyrights © 2026
