Why does fine-tuning on company docs often raise hallucination instead of lowering it?

Teams often fine-tune a model on their internal documents hoping it will learn the facts inside them. Surprisingly, the model frequently hallucinates MORE afterward, not less. The reason is what fine-tuning actually teaches. Supervised fine-tuning shows the model thousands of question-answer pairs, and the model learns the SHAPE of a good answer: the tone, the jargon, the confident structure. It does not get the trillions of tokens of repetition that pretraining used to actually memorise facts. So the model walks away sounding like a domain expert while knowing roughly the same facts as before. When you ask something it never truly learned, it answers in that confident expert voice anyway and makes things up. Confident plus wrong is the definition of a hallucination. The fix is to separate the two jobs. Use fine-tuning to set tone and format. Use **RAG** to inject the actual facts at query time, so the model reads the real document instead of guessing from memory.

Fine-tuning on company docs raises hallucination because of a mismatch between what the training signal teaches and what teams hope it teaches. The answer is A. Supervised fine-tuning on a few thousand examples efficiently teaches the style and confidence of domain answers. It does not reproduce the pretraining-scale memorisation needed to actually store the facts. The model learns to sound authoritative without learning the underlying knowledge. The loss also pushes the model toward sharp, confident output distributions. So on questions adjacent to or outside the fine-tuning set, the model now answers fluently about things it still does not know. Confidence rises while factual coverage stays flat, or even drops because of catastrophic forgetting. The correct tool for injecting facts is retrieval. Use fine-tuning for behavior, tone, and format. Use **RAG** to put the real document text in front of the model at query time, where it can ground its answer instead of confabulating from weak memory.