Thanks for comment. We don't train on the dataset being tested on. I am fairly certain the other engines don't as well.