almanach

/

moderncamembert-base

like 6

Follow

ALMAnaCH (Inria) 111

First, thanks for this amazing model!

I'm under the impression that in training stats the loss metrics are summed instead of averaged: contrary to camembert or camemberta, when I use a validation set of 20%, the reported trainset loss is 4x higher than the testset loss.

This is a bit misleading, see picture, especially wrt other models where the losses are averaged.

Thanks