About model training

[MoonLBH]

I want to retrain MIDI on the Chembl with explicit hydrogens. During the training process, I found that after more than 100 epochs of training, the validity of the model in the valid sampling is still relatively low, only about 0.2, but the mol stable is gradually improving. I am confused whether there is a deviation in the training or this is a normal trend. It’s just that the MIDI model needs to improve the mol stable first before improving the validity.