Adverse drug events (ADEs) automatic detection from social media posts has become an important task for healthcare systems with real-world, patient-collected data. The current work deals with ADE on user generated content for Task 1 of the Social Media Mining for Health Research and Applications Workshop (SMM4H 2026), Creative Catalysts. We fine-tuned XLM-RoBERTa, pre-trained model chosen for its robustness in handling multilingual content and linguistic diversity common in social media text. To better handle the class imbalance, we subsequently implemented a class-weighting strategy to increase the model’s focus on the underrepresented positive class. This adjusted model improved the validation F1-score to 65%. Our results demonstrate the effectiveness of transformer-based architectures for ADE detection while highlighting the critical need for robust class-balancing techniques and multilingual generalization to handle real-world, imbalanced social media data.