promptbase
promptbase copied to clipboard
microsoft
Clarification needed in evaluation numbers
Hello,
Thanks for the repo and awesome work. I am requesting clarification on the evaluation results shown in the repo.
For humanEval Zero shot, GPT-4's score is reported here as 87.4 but in the Gemini Report and GPT-4 paper(and everywhere else), humanEval score for GPT-4 Zero shot is 67.
Is the "Zero-shot" prompt technique mentioned in the repo followed by Medprompt methodology? If yes, please clarify. For MMLU is explicitly clear but not for others.
Apologies if I missed anything.
@saurabhkumar8112 look into their code, i guess it is standard zero-shot result using newest GPT-4 checkpoint.
Yes @dzunglt24 is right -- we do have all the code we used to run on HumanEval here, and it is zero-shot with the latest GPT-4 checkpoint. The numbers reported in the OpenAI report are from many months ago, and it's likely that there have been both model improvements, and subtlety in prompting differences (even in the zero shot setting) that leads to our improved performance number here.
I believe others have found that the GPT-4 numbers were underreported in the Technical Report as well, e.g. see: https://twitter.com/OwariDa/status/1732423557802782854
Our HumanEval scripts/prompt are: https://github.com/microsoft/promptbase/blob/f43cf97dd81c9595b7aec40a2201797b32532084/src/promptbase/humaneval/humaneval.py#L16
I see. That’s good to know. Then that means the Gemini Report had under-reported numbers for GPT4(as the numbers were from old checkpoint)?
I see. That’s good to know. Then that means the Gemini Report had under-reported numbers for GPT4(as the numbers were from old checkpoint)?
I believe the Gemini Report cited + pulled the Humaneval numbers directly from OpenAI's initial GPT-4 technical report (which was released in March alongside the first version of the model). We just happened to run our own zero-shot prompts against a more recent checkpoint so we have updated numbers here.