Prepare validation pipleine (notebook on kaggle at least with hold-out set)

[R1chrdson]

We should define hold-out dataset to evaluate the final CV score of ensembles. That hold-out set shouldn't be used on training