...It supports generation of rationales/chain-of-thought variants, configurable sampling, and guardrails so outputs meet format constraints and quality checks. Examples and guides show how to target task-specific behaviors like tool use or step-by-step reasoning, then save directly into training-ready files.