Train / validation split -- does it need to be left as-is or can we utilize the data as we like?

[maciej]

In particular, are we allowed to train on the whole (train, validation) dataset?