I guess it would mostly depend on training data uniqueness, yes?