ESM-IF1: how to reproduce result about GVP-Transformer trained with only CATH4.3

[ChangxinWang]

I want to reproduce the result about GVP-Transformer trained with only CATH4.3, but I can only find the training details about GVP-Transformer trained with (CATH+UR50) from Appendix Table A.1.

I tried to train GVP-Transformer with only CATH datasets, using the hyperparameters listed in Appendix Table A.1 (gvp-4layers + Transformer-8layers)，but I meet serious over-fitting, and resulting about recovery 20.0 on test split.

Could any hero give me more details about reproducing the result about GVP-Transformer trained with only CATH4.3? (just like the Appendix Table A.1). And I have another questions apart from hyperparameters:

1. does GVP-Transformer trained with Cross-Entropy ?