Catastrophic Forgetting
Catastrophic forgetting is the tendency of a neural network to lose previously learned knowledge when fine-tuned on new data, because weight updates overwrite representations learned during pretraining.
What is Catastrophic Forgetting?
Catastrophic forgetting is the tendency of a neural network to lose previously learned knowledge when fine-tuned on new data, because weight updates overwrite representations learned during pretraining.
Catastrophic forgetting is the tendency of a neural network to lose previously learned knowledge when fine-tuned on new data, because weight updates overwrite representations learned during pretraining.
Where is it used?
Fine-tuning Llama on a narrow domain without replay can degrade general QA ability; PEFT methods like LoRA and regularisation techniques like KL-divergence penalties (used in DPO) mitigate this.
How to build it
Fine-tune GPT-2 on a tiny code dataset, then evaluate perplexity on a general text corpus before and after; observe the perplexity increase indicating forgetting, then compare with a LoRA adapter.