Definition#
Catastrophic forgetting occurs because gradient updates that improve the target task also move weights that encoded other skills. In LLM fine-tuning it shows up as a model that extracts claims perfectly but can no longer follow an unusual instruction, write fluent prose or respect safety behaviours from its original training.
How to control it#
- Data mixing: include 5–20% general instruction data in the fine-tuning set.
- Parameter-efficient methods: Biderman et al. (2024) found LoRA forgets less than full fine-tuning, while also learning less on hard domains (comparison).
- Conservative optimization: lower learning rates, fewer epochs, KL penalties in preference and RL stages.
- Measurement: a general-capability regression suite run on every checkpoint, alongside the task evaluation.