TM-score for reconstruction

[lspindler2509]

When trying WrappedOurPretrainedTokenizer, we get TM-scores not above 0.5 when using the Casp15 multimer dataset. Is this as expected or is something going wrong?

[KatarinaYuan]

Hi, given our tokenizer is pretrained on single chains only, is would be important to check all the code logic when loading the multimer datasets, and it might be expected to perform badly on multimer datasets. Because I didn't try Casp15 before, I cannot say anything certain for the result. Have you tried to test monomers from Casp15?