Controls how strongly the model follows your prompts/conditions.
Low values: Model is more creative/random, High values: Model adheres strictly to prompts, less variation.

Nucleus Sampling: Limits the pool of possible next tokens to the most likely ones.
(Default behav.) only sample from tokens that make up the top 95% of probability

In masked models, controls the randomness of which tokens get unmasked first.
Low temperature: More deterministic, predictable unmasking order, High temperature: More random unmasking, can add variety

Same seed and same inputs: Produces identical outputs (reproducible).
Set a specific seed to get consistent results, or use fifferent seeds to try variations