Instead of benchmarking pre-trained models via their model identifier, e.g.