Is it possible to get the raw results for the pre-trained models for the three variants mentioned in the paper ?