A model trained to avoid harmful outputs and refuse unsafe requests, making it more cautious and responsible in its responses.
Adhering to complex, structured, or constrained instructions