Pierre Ablin

We propose scaling laws that predict the loss of models when trained on a mixture of source domains.