"Split out" validation dataset

[zeppelinche-cmd]

It would be good if we could split out a tiny part (e.g. 0.1%) of all of our source datasets early on in processing to serve as a validation dataset that we don't train on.

Relevant for post-flag efforts.