e.g. a model that is to GPT-4 as code-davinci-002 is to GPT-3.5. It does not need to be released by OpenAI - if for some reason Microsoft produces it, or the weights are leaked and a third party produces it that will also count, as long as it is clear that the base model is GPT-4. I also do not require that the base model be the exact GPT-4 version that is deployed to their API or that was used to produce the benchmarks in the technical report.
code-davinci-002 is base GPT-3.5. OpenAI seems pretty dead set on their mainline RLHF models being coding models.
What if they release a finetuning API?
@duck do you mean what if they release a fine-tuning API and someone fine tunes it for code then releases that model? That would resolve YES.
GitHub copilot X
@firstuserhere copilot X could just use base GPT4. GPT4 is already good at coding so they might not fine-tune it further, and they're already pulling code-davinci-002 off the OpenAI playground.