Why retrain that smaller model from scratch tho? Just do a little transfer learning, or get creative and see if you can prune down to a smaller model algorithmically instead of doing the whole label and train rigamarole from scratch on what is effectively regurgitation.
I’m not sold this has directional value.