This research focuses on the effectiveness of RoBERTa deep learning model in classifying depression based on social media posting patterns. Platform X provides a potential tool that can be used for detecting early signs of mental health disorders such as depression, through analysis of social media behavior. This research uses data from X users in Indonesia who have completed the DASS-42 questionnaire. The dataset contains 88,842 rows representing 138 active users, which include two labeled categories: positive and negative. This research explores two data preprocessing to compare model performance. The first approach, half preprocessing, involves case folding, cleansing, and tokenization, but excludes normalization, stopword removal, and stemming. The full preprocessing approach includes all steps of preprocessing techniques, including case folding, cleansing, tokenization, normalization, stopword removal, and stemming. The half preprocessing approach achieved highest accuracy at 58.07% with f1 score 57.14%. On the other hand, the full preprocessing approach resulted in a lower accuracy at 53.81% and f1 score 52.38%. The outcomes of this research underscore the importance of preprocessing approaches in optimizing the RoBERTa model’s performance. These results provide insight into refining data preparation strategies, which can lead to the development of more advanced models for mental health classification. By understanding the impact of different preprocessing techniques, future research may concentrate on developing models with enhanced capability to manage noise in the data.