There seem to be a couple of errors in the docs:
- They state that when logprobs are returned, they are for tokens, but this isn’t the case: they are for words (computed presumably from the constituent tokens)
- The definition of “best” in the documentation for
best_of
is wrong: it’s not “the one with the highest log probability per token” (whatever that could mean) but the one(s) with the highest overall probability(s) (sum of all its logprobs).