How to train a generative chatbot with new data without making it inaccurate for the previous data

Hi, I followed the PyTorch tutorial on creating a chatbot. I trained it on Cornell Movie-Dialogs Corpus. But what I need to do now is to append new data to the chatbot without losing previous weights. how can I do that?

here is the notebook for training the model (from a checkpoint):

And here is the notebook where I am loading the previous model and trying to add new data: