Comment by throwa356262

That's not benchmaxing.

You need a method to show you RL is improving the model that is separate from the training data.