My research relies extensively on evaluating the logprobs that an LM gives to different strings in different contexts, which works well for the Completions API but is not possible at all with the Chat API. Is there any plan to have a Completions API for a GPT-4-scale model?
4 Likes
From what I have learned (read in the manual at openai) is that GPT-4 is for both completions as chatgpt.
“GPT-4 is optimized for chat but works well for traditional completions tasks.”
Source: OpenAI API
I’m also curious about this as I have a similar use case. Any update on this? Can’t seem to find anything concrete about this.
Also curious to know the update here. Completion model has advantages like better control over the internal structure of a generation. This is currently not directly possible with a Chat model. It’s a pity that the best LLM out there (GPT-4) doesn’t support completion and people have to fallback to GPT-3.