Lesson 16: generation controls — the dials a developer tunes
Until now we improved the prompt itself — the words. But a developer also has a few 'dials' that change how the model generates the answer, without touching the wording: temperature (how creative or steady), max length (how long the answer is), and a stop sequence (where to stop). In this lesson we'
Like dials on a device: temperature is the creativity dial (low = steady and predictable, high = varied and surprising), max length caps the answer's length, and a stop sequence says 'stop here'.
- temperature
- The model's creativity dial: a low value gives steady, predictable replies; a high value gives more varied, surprising ones.
- max length
- A cap on the answer's length — the most the model is allowed to generate before it stops.
- stop sequence
- An agreed marker that, the moment the model produces it, makes it stop — a way to cut the answer at a precise point.
