Creating a better validation set when test examples differ from training examples - View it on GitHub