data inconsistency between pretraining model data and proteinnet

[zzhang-QSI]

Hi, alquraishi

I tried to train the model using CASP12 tfrecord data download from proteinnet: https://sharehost.hms.harvard.edu/sysbio/alquraishi/proteinnet/tfrecords/casp12.tar.gz
I found the training_90 data is different from pretrained model (CASP12)
RGN12/runs/CASP12/ProteinNet12Thinning90
https://sharehost.hms.harvard.edu/sysbio/alquraishi/rgn_models/RGN12.tar.gz

may I know which data is better?
Using proteinnet data , I found numEvoEntries change to 21 instead of 42, and the training loss always nan.

Looking forward to hearing from you.
Thank you