The domain gap
Sometimes a model follows instructions perfectly but still flounders, because it has barely seen your domain. Clinical shorthand, statutory citations, a proprietary codebase, or a low-resource language can all be underrepresented in the pretraining mix. No amount of prompt polishing teaches vocabulary the model never absorbed. This is the realm of domain adaptation.
Pipeline: a query retrieves documents that are added to the prompt before the model generates an answer.
Continued pretraining
The tool for closing a knowledge gap is continued pretraining (also called domain-adaptive pretraining): you keep going with the same self-supervised next-token objective the base model was built with, but now on a large corpus of your domain's raw text. There are no labels here — just lots of in-domain documents the model learns to predict. This is transfer learning at work: the general language skill carries over, and the domain settles on top.
Continued pretraining keeps optimizing the same self-supervised objective: predict each next token, minimizing the negative log-likelihood over your domain corpus.
A common high-end recipe is two stages: continued pretraining to install the domain, then supervised fine-tuning to restore instruction-following on top of it. Pretraining teaches what the world looks like in your domain; SFT teaches how to behave within it.
Catastrophic forgetting — the central danger
Push a model hard on narrow data and it can suffer catastrophic forgetting: as the weights move to fit your domain, they drift away from the general abilities that made the base model worth using. You end up with a model fluent in radiology reports that can no longer hold a normal conversation or do basic arithmetic. This is the failure mode that makes domain adaptation tricky.
- Mix in replay data — blend general-domain text back into the corpus so old skills keep getting reinforced. Tuning this mixture is a data-mixture-weighting decision.
- Use a low learning rate and few epochs — gentle nudges preserve more of the original model than aggressive runs.
- Prefer PEFT when you can — adapters confine most changes to a small add-on, leaving the frozen base intact and far less prone to forgetting.
Adaptation vs the knowledge cutoff
Domain adaptation is also one way to move a model past its knowledge cutoff for a specific area — continued pretraining on recent in-domain text bakes newer facts into the weights. But baked-in facts go stale the moment the world changes again, and retraining is slow. For information that updates frequently, retrieval is usually the better answer; reserve adaptation for durable domain knowledge and idiom.
A domain-adaptation recipe
- Confirm the gap is really domain knowledge, not prompt or format — and that retrieval can't cover it more cheaply.
- Gather a large, clean corpus of in-domain text and a slice of general text for replay.
- Continued-pretrain at a low learning rate for a few epochs, watching general benchmarks for signs of forgetting.
- Follow with supervised fine-tuning to restore instruction-following and your target response style.
- Evaluate on both a domain test set and a general one, so an in-domain win can't hide a general loss.