loss functions for fine-tuning embedding and reranker models

[hank317]

May I ask whether the BGE-M3-embedding and BGE-reranker-MiniCPM models use InfoNCE loss during fine-tuning?

From the code implementation, it appears that cross-entropy loss is used. Can I interpret this as using a cross-entropy formulation while still being essentially equivalent to InfoNCE in principle?

[545999961]

Yes, you are right, this is equivalent to the InfoNCE loss.