Xin1212 at NakbaVirality Shared Task: Frozen CLIP with Residual Adapter for Multimodal Virality Classification

Proceedings of the 2nd International Workshop on Nakba Narratives as Language Resources @ LREC 2026

Abstract

We describe our system for the NakbaVirality shared task on multimodal virality classification. Our final approach uses a frozen LAION CLIP backbone, a lightweight residual adapter over fused text–image embeddings, and a small MLP classification head. On our development split, the best configuration (V9) achieves Macro-F1 of 0.5492 and virality-weighted F1 of 0.5252. On the official test submission, our system obtains F1-score 0.4559 and accuracy 0.6089 according to the platform scorer. We provide implementation details, ablations across multiple versions (Baseline–V9), and practical error analysis for reproducibility.