Temperature
A divisor applied to the logits before the softmax, setting how much probability mass concentrates on the top candidates. Below 1 sharpens towards the likeliest token; above 1 flattens and lets unlikely ones through.
What people get wrong
Temperature 0 is not a real temperature — dividing by zero is undefined. It is the libraries' name for switching to greedy decoding.
A definition is not an understanding. This is taught properly in How a model picks the next token, part of Sampling & decoding.
Learn this properlyTerms from the same topic