A community pageant troupe wants the model to follow commands such as “write a 30-second curtain speech.” The training set is instruction plus desired speech pairs. Which method is that?

Instruction tuning trains on instruction-plus-desired-response pairs so the model reliably follows directive-style prompts such as “write a 30-second curtain speech.” That matches the community pageant troupe’s labeled set of command and desired-speech pairs and is why instruction tuning is the right method name. Continued pre-training on unlabeled diaries improves language familiarity but does not teach follow-the-command behavior for curtain speeches. Prompt caching of a rainfall table reuses a static prefix and is unrelated to instruction-following training. A ticket-sales dashboard is analytics for last year’s attendance, not a customization method that teaches the model to obey stage directions. Exam caveat: instruction-tuned models still need grounding and safety controls; following instructions well includes following harmful ones if those are not filtered at the application layer. Operational check: evaluate a held-out set of curtain-speech instructions for format, length, and tone compliance after instruction tuning before handing the model to volunteer stage managers. Refresh the instruction pairs when the troupe changes speech length or tone standards so the tuned behavior stays aligned with live shows.