Why is deberta-v3-large model twice as large on disk after MLM finetuning? (notebook to reproduce)

On the model hub large DeBERTaV3 is about 800mb:

But training with run_mlm.py for even just a few steps results in a model that is 1.6gb:

Any idea why this is happening and if it’s possible to reduce the size of the fine-tuned model?