The explosive development of social networks has revolutionized the way people interact, but it has also enabled the emergence of cyberbullying, a widespread phenomenon with severe psychological and social implications. Conventional moderation techniques are usually inadequate in dealing with this problem because they are not scalable and flexible. To address this, we systematically evaluate both traditional machine learning (ML) methods and deep learning (DL) methods. This study utilizes two datasets: a balanced multiclass dataset and a second dataset from the study by Davidson et al. (hereafter referred to as Davidson). A major emphasis is placed on the systematic evaluation of performance across datasets to understand their effectiveness in detecting harmful content. Experimental results show that the RoBERTa model performs better on the balanced multiclass dataset with an accuracy of 0.8604, while BERTweet achieves the highest accuracy of 0.9207 on the Davidson dataset. In general, our study systematically evaluates several ML and DL methods on two different datasets.