Merging two classification models

I have a lot of labelled data for two different labels and trained a model for each one of those. I.e. I now have a model that can predict class A and a model that can predict class B. The training took a couple of hours for each class.

I'd like to combine the two models into a single model that predict. A few methods to proceed:

  1. I could of course combine my data and retrain on all of it but I prefer not two, since I'm going to repeat this exercise for multiple labels over time.
  2. Update model A to learn about class B and making sure it won't do any catastrophic forgetting about A. Is there a good recipe for this?
  3. Other options?

I could imagine a lot of people have come across this task so maybe someone has a great approach. Thank you in advance!

If the training only took a couple of hours, I'd definitely recommend 1 over 2. It's just much easier to reason about if you're always training from scratch. I'll put it this way: if you train with approach 2, you're going to get some accuracy back, and then think to yourself "I wonder whether it would be more accurate if I trained it from scratch". So you'll have to train with approach 1 anyway.