Clarifications on Log Probabilities for Chat Completion

[cyzgab]

Firstly, are the log probabilities natural logs or in base 10? Could not seem to confirm this in the documentation.

Secondly, curious about the design - why does the endpoint return the log of the probability, and not the probability itself?

[_j]

2. Because the values of less-likely tokens can be very small, returning the exponents like -9.3 is possibly seen as better than .00000003, and they are in that form for temperature scaling and other unknowns.

Answer 1 and a bit of code for adding normal probabilities to the return object or stream, that I haven’t yet adapted to new API, but probably would work on a response.model_dump() of the return object.

Logprob value is unbounded, i used sigmoid (converting API logarithm to probabilities)

The logprobs are simply that, probabilities, represented logarithmically (base e). I think what you wish for is if the API returned the probabilities as they are shown in the playground, and then also gave the total probability coverage each token. We can do that. some imports import os import openai import json import math import copy A function to accept a completion return, and if it has logprobs, add a new item with the same logprob structure, but with 0-1 probabilities instead of exponen…