Fine-tuned DistilBERT (67M parameters) for hate speech detection with attention-based token attribution — highlights which words most influenced the classification.