Did WizardCoder generate data from GPT-4 or GPT-3.5?

[Symbolk]

Thanks for this repo! I am also reading the paper recently, but I did not notice which LLM the WizardCoder used to generated their Evol-Instruct data. According to your implementation, gpt4_azure is used, is it the same with WizardCoder (considering that Microsoft insiders could use the API for free since early this year), or you just guess they used GPT-4?

[swtheing]

Yeah, the paper show the instructions are generated by GPT-4 or GPT-3.5. But, how to generate the response of the new generated instruction is not explicitly shown in their paper. I guess the response should be generated by GPT-4

[Symbolk]

Yeah, the paper show the instructions are generated by GPT-4 or GPT-3.5. But, how to generate the response of the new generated instruction is not explicitly shown in their paper. I guess the response should be generated by GPT-4

Do you mean only the ###Response is generated by GPT-4? A third choice is generating programming tasks with GPT-3.5, and then pass the task to GPT-4 to generate the solution.

FIY: As the just leaked GPT-4 detail shows (https://threadreaderapp.com/thread/1678545170508267522.html), GPT-4 is trained 2 epochs for text-based data and 4 for code-based data!

[jaideep11061982]

@Symbolk my objective is to do just batch inferencing in that case what is going to be my prompt template with an eg. if you can explain thanks in advance

[haorannlp]

Hi, did you successfully reproduce the training data?