threat-classification

SMETA-ZSL proposes a method for generalized zero-shot threat classification using semantic meta-alignment and contrastive finetuning, outperforming prior methods by 10.8 points on average across 7 benchmarks.