Catastrophic forgetting during fine-tuning

Catastrophic forgetting is the tendency of a fine-tuned model to lose previously learned capabilities (general reasoning, instruction following, language coverage) when trained on a new narrow dataset. Mitigations: (1) replay: mix in 10-30% general-domain data during fine-tuning; (2) LoRA: keeping base weights frozen means no catastrophic drift; (3) KL-regularization to the reference model (used in RLHF/PPO and applied to SFT as "anchor loss"); (4) PEFT methods like IA3, prompt tuning learn tiny parameter sets that perturb behavior minimally; (5) multi-task / instruction-mix fine-tuning (Tulu, FLAN): never single-domain SFT; (6) staged curricula: world-knowledge SFT followed by capability SFT followed by domain SFT. Senior nuance: empirical reality in 2025-2026 is that a strong SFT-then-DPO pipeline on well-curated open data is usually less destructive than narrow-domain SFT alone.