Open the workspace
Glossary

Catastrophic forgetting

Catastrophic forgetting is the loss of general capabilities after fine-tuning on a narrow dataset. How to measure it and the mitigations that work: data mixing, adapters, KL and early stopping.

POFE ResearchPublished 2026-09-05Updated 2026-10-304 min read

Catastrophic forgetting: The degradation of a model's previously learned abilities after training on new, narrower data.

Explained#

Fine-tuning moves weights toward the new data distribution; abilities the new data doesn't exercise can degrade. In LLMs it shows up as worse instruction following outside the target task, lost safety behaviour, broken formats, or reduced reasoning.

Measuring it

Keep a small "regression suite" of general tasks — instruction following, refusals, a reasoning set, your own previously shipped tasks — and run it on every candidate alongside the target eval. Forgetting you don't measure, you ship.

Mitigations

  • Data mixing (rehearsal): include 5–20% general instruction data in SFT.
  • Adapters: LoRA forgets less than full fine-tuning at the cost of learning less (Biderman et al., 2024).
  • KL to the reference in preference and RL stages.
  • Fewer epochs, lower learning rate, and stopping when the target metric plateaus.

References#

  1. Biderman, D. et al. (2024). LoRA Learns Less and Forgets Less. arXiv:2405.09673. arxiv.org/abs/2405.09673
LoRA (Low-Rank Adaptation)GlossaryKL divergence penaltyGlossaryWhat is post-training? A complete guide for LLM teamsPillar guideLoRA vs full fine-tuning: quality, cost and serving trade-offsGuide

Bring a model and a metric.

Tell us what the model should get better at. A post-training engineer replies within one business day with a plan, a compute estimate and an eval design.