Don’t Judge a Language Model by Its Last Layer: Contrastive Learning with Layer-Wise Attention Pooling

4819

Knowledge Triples