Antiviral peptides (AVPs) hold significant potential in public health by effectively inhibiting a variety of viruses. However, existing research primarily relies on traditional models to predict their activity, which are limited by feature extraction capabilities and model generalization. To address these limitations, we propose a two-stage predictive framework that integrates the ESM2 model, data augmentation, feature fusion, and contrastive learning techniques. Utilizing an innovative top‑k queue-based contrastive learning strategy, this framework significantly enhances the accuracy and generalization performance in identifying AVPs and their subclasses. The model demonstrates excellent performance on independent datasets, showcasing the strong application potential of this method in antiviral peptide research.