efficient way of saving finetuned zero-shot models?

Hi guys, pretty interesting project.

I was wondering if there is any way to save models after a zero-shot model is finetuned for few-shot model.

So for example, if I finetuned a couple of say, `sentence-transformers/paraphrase-mpnet-base-v2` models, the major difference between them is just the weights of final few layers, weights for the rest of the model mostly remains the same, so is there a way to efficiently save the necessary final few layers thus reducing the size of models, repetedly being saved.

This way one could save, the disk space by a lot.

And apart form that, while inferencing, I don't have to load multiple huge models and instead I could have just one model containg the common freezed layers that give me some common features and just has to host the final few layers with custom classes that intakes those common features.