💭

• really important you dont train on padded indices
• you can't take ml-dependent data processing out of the train cycle for performance or it doesnt train silly