Fine-tuning helps the model use this knowledge more efficiently, but introducing new knowledge through fine-tuning can increase the risk of hallucinations.