NEURAL_TRAIN is called at the end of any loop() cycle. This is necessary for training different assets and algos.

NEURAL_SAVE and NEURAL_LOAD are of course called at the end of a WFO cycle, since they are supposed to store and load all the trained models. Saving them under a name like SCRIPT_NAME_ASSET_1.ml makes no sense, since there is no particular asset. I think the support got it wrong here. Use the file name given by the Data string and save not only one, but all previously trained models in that file.