A model must follow a specific response style, but the business facts are supplied separately at inference time. Which customization objective is most appropriate?

1. Evaluate supported fine-tuning on curated examples of the desired behavior and style.
2. Remove all current evidence because style tuning guarantees factual freshness.
3. Use unreviewed failed responses as correct training targets.
4. Choose the noisiest sampling settings without evaluation.