Your fine-tuned LLM forgot its general capabilities (catastrophic forgetting). How do you fix it?
LLM & AILevel: IntermediateTroubleshooting Scenarios
Your fine-tuned LLM forgot its general capabilities (catastrophic forgetting). How do you fix it?
Hints
- Which parameters change and which stay frozen?
- Separate memory for weights, gradients, optimizer states and activations.
Answer
- Switch to LoRA — if you used full fine-tuning, the primary fix is using LoRA; frozen base weights cannot forget
- Add data mixing — re-run fine-tuning with 10–20% general-domain data mixed in
- Reduce training intensity — fewer epochs, lower learning rate, reduce LoRA rank
- Merge carefully — if merging LoRA into base, use a lower scaling factor (α/r ratio)
- Evaluate incrementally — track both task metric and general benchmarks during training; apply early stopping when general capability starts degrading
Source and licence: JustInCache & contributors — Awesome AI Interviews · Apache-2.0