<![CDATA[Social media platform X has become a key channel for expressing public opinion on political issues, including evaluating the early performance of the government. The first 100 days of an administration are a strategic period to assess policy direction and public perception. This study aims to apply and evaluate the IndoBERT model for sentiment analysis of Indonesian-language tweets discussing the 100-day performance of the Prabowo–Gibran administration, as well as to assess the impact of using the Synthetic Minority Oversampling Technique (SMOTE) to address data imbalance. A total of 15,027 tweets were collected through API crawling and processed through several stages: preprocessing, labeling using the InSet Lexicon, data splitting, and fine-tuning IndoBERT. Two scenarios were tested — without SMOTE and with SMOTE oversampling. The results show that both models achieved the same overall accuracy of 87%, but performance varied across sentiment classes. The model without SMOTE performed better in the positive class with 93% precision, whereas the SMOTE-applied model improved performance in the neutral class (F1-score increased from 70% to 71%; recall from 69% to 71%) and in the negative class (precision increased from 88% to 90%). Considering the balance across classes, the SMOTE-based model was selected as the final model and implemented into a Streamlit application for interactive sentiment analysis. This study expands the application of IndoBERT in the Indonesian political domain by combining the lexical InSet approach with SMOTE oversampling — a combination rarely applied in Indonesian political sentiment analysis. The findings highlight the importance of data balancing strategies in improving transformer-based model performance on imbalanced datasets. Future research is encouraged to explore alternative balancing methods, expand training data, and test other transformer variants to enhance accuracy and generalization.]]>
