Catastrophic forgetting
Beginner
When learning a new job makes a model worse at things it used to do well.
Novice
The loss of earlier abilities when a model is trained on new data. A model fine-tuned for one task can get worse at general knowledge, math or following safety rules.
Expert
Degradation on the source distribution after fine-tuning, observed across 1B–7B LLMs. LoRA forgets less than full fine-tuning; mixing general data and evaluating held-out abilities are the standard guards.
See also: LoRA (low-rank adaptation), Fine-tuning.