I think it's commonly accepted that kfold cross-validation should be done with k=5 or k=10. We do 3 fold by default. Actually in a recent talk Max Kuhn said "never do threefold" and I think he was not even aware that this could have been a jibe.

Again, not sure if we should tag this 1.0 and make it a breaking change or just deprecate it because really no-one should have been using the default anyway.

If someone wants to see data, I'm not opposed to collecting data on this. Actually I think @GaelVaroquaux has some pretty in-depth studies of cross-validation, maybe he even has data.