What is catastrophic forgetting and how do you limit it?

Catastrophic forgetting is when fine-tuning on a new, narrow task makes the model worse at things it used to do. The gradients from the new data move shared weights without any signal to preserve other skills. A model tuned hard on medical summaries, for example, might become worse at following general instructions, or lose some of its safety behavior.

I limit it in a few ways. Parameter-efficient fine-tuning such as LoRA trains small adapter weights while the base weights stay frozen, which reduces drift. I mix a portion of general instruction data, or examples of the original behaviors, into the training set so the model keeps practicing them. And I train conservatively, with a modest learning rate, a small number of epochs and early stopping based on a validation set.

Most importantly, I measure it. Before and after training, I run the task eval plus a regression suite that covers general instruction following, reasoning and safety refusals. A fine-tune that improves the target task by a few points but breaks refusals is not shippable.