Catastrophic forgetting
Catastrophic forgetting happens when fine-tuning improves a narrow target behavior but damages capabilities the model already had. The model did not "forget" like a person; its weights shifted away from useful general behavior.
Why it happens
Fine-tuning data is usually much narrower than pretraining data. If training pushes too hard on that narrow distribution, the model over-specializes.
| Cause | Effect |
|---|---|
| High learning rate | Large, destructive weight updates |
| Too many epochs | Memorization and drift |
| Narrow dataset | Loss of general behavior outside the target task |
| Bad formatting | Chat/template behavior regresses |
How to detect it
Evaluate both the target task and preserved behaviors:
- Held-out task examples from the fine-tuning domain.
- General instruction-following prompts.
- Safety/refusal tests if policy behavior matters.
- Format and tool-call tests if the model is used in a system.
- A small benchmark or smoke suite for broad capability regression.
Mitigations
Use lower learning rates, early stopping, smaller LoRA rank, regularization, and a mixed dataset that includes examples of behavior you must preserve. Keep a reference model and compare outputs during evaluation.
Pitfall
If you only evaluate the target task, forgetting looks like success. Always measure what must not get worse.
Connects to: LoRA · eval suite · overfitting