Model Parameters

`options`, one dict for everything

Every call so far has used a model's defaults. ollama.chat() and ollama.generate() both accept an options dict that tunes how the model generates, separate from what you ask it. This lesson covers the three you'll reach for most: temperature, seed, and num_predict.

`temperature`: randomness

At each step, a model doesn't pick one "correct" next word, it computes a probability for every word in its vocabulary and samples from that distribution. temperature controls how sharply that sampling favors the most likely option:

• temperature=0.0: always take the single most probable next word. Deterministic (same input, same output) and reliable, but repetitive, the model can't surprise you or vary its phrasing.
• Higher values (1.0 is a common default, this lesson goes to 1.5): let less-likely words get picked sometimes, producing more varied, more "creative" output, at the cost of losing that repeatability and, past a point, coherence.

There's no universally right value, low temperature for factual lookups and structured extraction (Lesson 10), higher for brainstorming or creative writing.

`seed`: reproducibility

Even at temperature=0.0, some randomness can creep in from how the model's internals are computed. seed pins the random number generator itself to a fixed starting point, so the same prompt with the same temperature and seed produces the identical answer every time, useful for tests and debugging, where "did my prompt change actually change the output" needs a real answer.

`num_predict`: a hard length cap

temperature shapes what the model says; num_predict limits how much. It's a hard ceiling on the number of tokens generated, regardless of whether the model was in the middle of a sentence. Set it low enough and you'll see the answer cut off mid-thought, which is exactly what the third example below demonstrates on purpose.