When training, you should usually set 
add_generation_prompt=False, because the added tokens to prompt an assistant response will not be helpful during 
training.