How do you mitigate catastrophic forgetting when fine-tuning a foundation model on a highly specialized, narrow dataset?
