InfoNCE Contrastive Loss

Embedding models (CLIP, sentence-transformers, E5) are trained contrastively: in a batch of matching pairs $(a_i, b_i)$, each $a_i$ should be closer to its own $b_i$ than to every other $b_j$.