The scikit-learn folks are planning to change the default number of CV splits for all kinds of cross-validation splitters from 3 to 5 since 3 has too much variance.

Should we also do this in for our default grid_search_fold and SKLL_MAX_CONCURRENT_PROCESSES values? This would mean our SGE workflows would change to require 5 slots instead of 3 whenever we run on the grid. However, as we have also seen, variance among the folds is bad.