Each process picked its own train/test set, which is bad