Что такое Temperature?

Temperature is an LLM inference parameter. At temperature 0, the model picks the highest-probability next token every time — output is deterministic and (mostly) reproducible. At higher temperatures, the model samples from a wider distribution — output is more varied and creative. Practical guidance: use 0-0.3 for factual Q&A, code generation, classification (you want consistency); use 0.6-0.8 for creative writing, brainstorming, summarization with style; rarely go above 1.0. Combined with `top_p` (nucleus sampling) for finer control. Temperature is the single most under-tuned parameter — many teams ship at the default (often 1.0) when 0.3 would give better results for their use case.