Social media platforms generate huge volumes of user-generated content on daily basis, which often includes toxic comments that highlight online conflicts in the form of social fandom war. These conflicts pose active challenges for digital content moderation. With the enhancements of natural language processing, researchers have developed advanced methods to detect and interpret toxic behavior in online communication. Deep learning has played a significant role in this progress, empowering models to capture complex and semantic patterns that traditional rule-based or shallow machine learning approaches are unable to detect them accurately. The main aim of this study is to integrate advanced large language models, Yi-34B characterized by bidirectional attention mechanisms and dual-stage training to enhance the classification of toxic social media comments. The approach combines both traditional textual features such as term frequency-inverse document frequency, parts-of-speech tagging, and bigrams with deep semantic features including advanced sentence embedding and topic modeling using Latent Dirichlet Allocation method. Empirical analysis demonstrates that the proposed model achieves the highest accuracy of 98%, compared to BERT and RoBERTa. Furthermore, explainable AI techniques are used to interpret model decisions and transparency based on global (SHAP) and Local (LIME) features explanation. In addition, statistical significance tests validate the model’s accuracy and per-class performance, which ensure its state-of-the-art effectiveness in social media toxicity classification.