> All this despite Italian accounting for marginal percentage of the training set.

Evidence?