Cappy: Outperforming and boosting large multi-task language models with a small scorer

Cappy is a small scorer that boosts the performance of large multi-task language models. It functions either independently on classification tasks or as an auxiliary component for generation tasks. Cappy can adapt multi-task LLMs without requiring back-propagation through LLM parameters, reducing memory requirements. It outperforms other LLMs and achieves parameter efficiency.