Contents

temperature

Temperature influences the confidence of the models response.

Declaration

var temperature: Double?

Discussion

The value of this property must be a number between 0 and 1 inclusive.

Temperature is an adjustment applied to the probability distribution prior to sampling. A value of 1 results in no adjustment. Values less than 1 will make the probability distribution sharper, with already likely tokens becoming even more likely.

The net effect is that low temperatures manifest as more stable and predictable responses, while high temperatures give the model more creative license.