Temperature

Temperature is a numeric parameter that determines how “random” the language model chooses the next word. When temperature=0, the model always selects the most likely token - the responses are repeatable and predictable. When temperature = 1-2, the model more often selects less probable words - the answers are more varied, but there are also more errors.

In my practice, temperature=0 or close to it - for tasks where accuracy is needed: data parsing, structured extraction, classification. Temperature=0.7–0.9 - for copywriting and generating variants: varied enough not to be boring, coherent enough to publish. Temperature>1 - only for experiments, the output is usually nonsense.

It is important to understand: temperature is not “creativity”. This is the degree of randomness in sampling. True “creativity” comes from the training data and model architecture. Raising the temperature above 1.2 when working with the API rarely makes sense - usually it is enough to write in the prompt “suggest 5 non-standard options”.