The rapid growth of digital platforms like YouTube has generated massive volumes of unstructured user feedback, presenting significant challenges for manual analysis due to the prevalence of informal language and extreme class imbalances in sentiment distribution. This study addresses these issues by developing a robust sentiment classification system for Indonesian YouTube comments within the automotive education domain, leveraging the IndoBERT pre-trained transformer model to effectively handle linguistic variability and data disparity. The research methodology employs an IndoBERT-base model fine-tuned on a labeled dataset of 9,844 comments, integrating a comprehensive preprocessing pipeline for slang normalization and a Random Over Sampling (ROS) strategy to mitigate the bias toward majority classes. Experimental findings demonstrate that the proposed model achieves a commendable overall accuracy of 80.99%, with specific F1-scores of 0.84 for positive, 0.73 for neutral, and 0.72 for negative sentiments, significantly outperforming traditional baseline methods. Notably, the application of ROS successfully improved the recall rate for the minority negative class from 45% to an impressive 68%, ensuring better sensitivity to critical feedback. In conclusion, the integration of IndoBERT with targeted data optimization techniques offers a resilient solution for analyzing public opinion on social media, proving that transformer-based architectures can overcome the complexities of under-resourced languages and imbalanced data distributions to provide actionable insights for content creators.