A 2-point jump on 200 questions is usually noise. How to compute error bars, run paired comparisons and size your eval set to detect real gains.
