that is for sure what everyone does. also they train on evals with the datasets that they would be bench against.
(The loose version of this that’s true is that there may exist eval data contamination in pretraining. This is a hard problem to fully solve.)