I noticed that the current loss function is passed to prepare_model() via Config.train_loss, and the default is set to MSE.