Enhancing Input-Label Mapping in In-Context Learning with Contrastive Decoding