The set of molecules used for training is in "5cls_rmsaltol.csv". If you try to predict from an image of any of these molecules the model will almost certainly be correct because this was the training data. 5cls_single_novalidation.model.h5 uses only the molecule with index 0 from 5cls_rmsaltol.csv. 5cls_single_valset0.h5 and 5cls_single_valset1.h5 both used 20% of the data as validation set - the respective indexes of those validation sets are givent in 5cls_newskf0.csv and 5cls_newskf1.csv