Hi! It seems like your validation does not use any cropping (neither token-level cropping nor atom-level truncation), yet seem to get good validation LDDT scores early on even when training at 384 tokens. What do you attribute this generalization capability to?
I realized your training crops only token-level but leaves atom length unbounded. Could this be why?
FYI in contrast other models like Boltz crop both token and atom level at both training and validation
Hi! It seems like your validation does not use any cropping (neither token-level cropping nor atom-level truncation), yet seem to get good validation LDDT scores early on even when training at 384 tokens. What do you attribute this generalization capability to?
I realized your training crops only token-level but leaves atom length unbounded. Could this be why?
FYI in contrast other models like Boltz crop both token and atom level at both training and validation