First cut down the number of trials then create repetitions of train/test splits, like done now.
Or cut down the down the number of trials individually for each repetition, more complicated to implement as we can not use n_splits of StratifiedShuffleSplit
Add a parameter that limits the number of trials used in decoding.