Do you want to see scaling curves wrt data and param size? I agree that 1.2B and 10B tokens is not representative, but what scale of parameters and dataset sizes would be convincing?