The GPT-3.0 "davinci-instruct-beta" models have been returning non-deterministic logprobs as early as early 2021. This is speculation. CUDA itself often has non

The GPT-3.0 "davinci-instruct-beta" models have been returning non-deterministic logprobs as early as early 2021. This is speculation. CUDA itself often has nondeterminism bugs. text-davinci-001 and text-davinci-002 were trained through FeedMe and SFT, while text-davinci-003 was RLHF; the models themselves have more variance at high temperature.