Why are my trained weights 5.0GB, while the pre-trained weights are only 1.7GB?

I trained the GeoLayoutLM model and noticed that the weights I trained are 5.0GB in size, while the pre-trained weights are only 1.7GB. The large size of my trained weights is problematic for storage and deployment purposes, and I would like to explore ways to reduce the size without losing performance. Could anyone explain why there is such a large difference and suggest possible ways to reduce the size of my trained weights?

Thanks!