tool News
The set that controls how random or predictable will behave when producing the model output.
Temperature parameter (temperature) is a sampling setting that controls how much "skin" or "soft" the probability distribution when choosing the next word of a language model. The model calculates a possibility for each possible word in the vocabulary in each step (with the softwaremax function); the temperature parameter divides logit values before calculation. When the temperature is lower than 1, high probability words become more dominant, and the model almost always chooses the most possible word; this produces consistent but uniform outputs. As the temperature outputs over 1, the probability distribution becomes straightened, the chance to choose low-angled words increases, and the output becomes more diverse, creative but more predictable and error-opened.
This parameter is set according to the use scenario: the code production, mathematical solutions, or the tasks requiring certainty such as the postal question-cevap are preferred at low temperature (0 to 0.3), while the story typing can give higher temperature (from 0.7 to 1.2) more interesting results in creative tasks such as brainstorm or poem. When the temperature is set to zero, the model becomes decisiveistic and always chooses the most possible word; in practice many APIs offer this parameter as a direct adjustable option to developers.
This parameter is set according to the use scenario: the code production, mathematical solutions, or the tasks requiring certainty such as the postal question-cevap are preferred at low temperature (0 to 0.3), while the story typing can give higher temperature (from 0.7 to 1.2) more interesting results in creative tasks such as brainstorm or poem. When the temperature is set to zero, the model becomes decisiveistic and always chooses the most possible word; in practice many APIs offer this parameter as a direct adjustable option to developers.
Interactive experiment preparing...