Also known as: catastrophic interference, forgetting
When fine-tuning a pre-trained model on a new task erases capabilities the base model originally had. The classical neural-network failure mode that dominates fine-tuning practice — and the reason LoRA, mixed-data training, and rehearsal exist.
Catastrophic forgetting is the failure mode where fine-tuning a pre-trained model on task A causes it to lose capability on task B that it previously had. The phenomenon was named in the 1980s and has been a thorn in transfer learning ever since — neural networks have no inductive bias toward preserving knowledge they’re not currently being optimized on.
Standard gradient descent moves weights to minimize loss on the current data distribution. If a weight that was important for solving task B is also useful (or just slightly useful) for solving task A, the optimizer will adjust it to fit task A — possibly destroying its task-B function. There’s no explicit penalty for this in the vanilla loss; the only “memory” the network has of task B is whatever happens to be encoded in weights the optimizer doesn’t touch.
For LLMs, “task B” is everything — general world knowledge, language fluency, reasoning, code, math. A model fine-tuned aggressively on legal documents can lose the ability to write Python; a model fine-tuned on classification labels can lose the ability to follow open-ended instructions.
Custom-trained rerankers hit a milder version: a reranker fine-tuned hard on legal documents loses general retrieval quality on out-of-domain queries. The fix is the same playbook — train via LoRA on the base reranker, mix in a fraction of the original training data, validate on out-of-domain held-out sets.
Every fine-tune trades some general capability for some specific capability. The job is to make the trade favorably.
The mechanical reason is that LoRA literally doesn’t touch the base weights. The frozen weights
This is fundamentally different from full fine-tuning, where every weight is a candidate for the optimizer to overwrite. With full fine-tuning, the only way to preserve a capability is for the gradient with respect to your task loss to happen to leave the relevant weights alone, which it generally won’t.
The trade-off is expressivity. LoRA can only represent perturbations of rank
The practical guideline: try LoRA first. If LoRA can’t reach your quality target, then full fine-tune with rehearsal. Full fine-tune without rehearsal is almost always wrong outside research contexts.
Rehearsal mixes a fraction (typically 5-20%) of the original pretraining distribution back into the fine-tuning data. The fine-tuning loss now has two components: the new task and the old distribution. The optimizer can’t make the new task loss go down by overwriting the old-distribution capability, because the old-distribution loss would go up.
The effect is that gradients pull weights toward solutions that perform well on both distributions. There may be no such solution that’s also optimal on the new task — the model has to find a compromise. The compromise is usually much better than the unmitigated forgetting that full fine-tuning produces, but it’s slower to converge on the new task and may plateau at a slightly lower task-only score.
The exact ratio matters less than people think. The literature finds 5-20% rehearsal data works for most cases; below 5% the regularization is too weak; above 20% you’re effectively just adding more pretraining and the new task converges slowly. The harder question is what to rehearse — the original pretraining data, a held-out general-instruction dataset, or a domain-specific anchor set. The right answer depends on which capabilities you most need to preserve.
The strategic framing: rehearsal isn’t a hack, it’s a regularizer. It adds a prior that “the model should still perform well on this distribution” and lets the optimizer balance that prior against the task loss. Like all regularizers, it’s a knob — and like all knobs, the right setting is empirical.
Gradient descent on a narrow distribution moves weights toward solutions that minimize loss on that distribution — including by overwriting weights that encoded other capabilities. Without explicit pressure to preserve prior knowledge, the optimizer has no reason to leave it intact.
LoRA freezes the base model and trains only a tiny low-rank adapter. The original weights — and everything they encode — are physically unchanged. The adapter can only steer behavior on top of, not overwrite, the base capability. This is the cleanest mitigation in practice.
Mixing a fraction (typically 5-20%) of the original pre-training distribution into the fine-tuning data. The optimizer is forced to keep performing on the original distribution, which preserves the underlying capabilities. Costs: dataset prep complexity, slightly slower task convergence.
