Choosing Settings for the Task

Settings strategy

Part of: Temperature & Sampling

You now have three dials, temperature, top-p, top-k. The beginner mistake is fiddling with all of them at once. The pro move is simpler: start from what the task needs, change one thing, and usually leave top-p alone. This lesson is the cheat sheet. What it is Picking settings is just matching randomness to the goal. Two questions sort almost every task: 1. Is there one right answer, or many good ones? One right answer (extract a date, classify sentiment, fix a bug) wants low randomness. Many good ones (slogans, story ideas, alt phrasings) wants higher randomness. 2. Do you need the same output every time? If yes, for tests, caching, reproducibility, push toward determinism. How it works Here's the working playbook. In practice you mostly move temperature and leave top-p ≈ 0.9 . A few rules that keep you out of trouble: - Move temperature first. It's the biggest, most intuitive lever. Only touch top-p/top-k if you specifically want to cap the candidate pool. - Don't crank both temperature and top-p high together. That stacks looseness on looseness and tips into incoherence. If temperature is high, a tighter top-p (~0.9) keeps it grounded. - For anything verifiable, default low. Fac

Challenge: The Settings Advisor