These labels
should be the expected prediction of the model: it will use the standard loss in order to compute the loss between its
predictions and the expected value (the label).