How would you fine-tune a model for a specific security use case while maintaining safety guardrails?
