Yes I'm very much talking about data used in the RL section, I'm sure Google are indeed very good at pre-training.