Cursor sessions are pretty much what composer models are RL'd on. This bench and the training data are/should be basically the same distribution.