catastrophic forgetting

[sailfish009]

There seems to be catastrophic forgetting after model fine-tuning training.
How do you usually resolve this? The model's performance on the downstream task has improved.
However, realize that the model's ability before training has been compromised. Should retrain it?