Using J-K fold Cross Validation to Reduce Variance When Tuning NLP Models