[Discussion] Adding more diverse annotators representing subpopulations? about alpaca_farm

[lxuechen]

Thanks for your interest!

I plan to study diversity-related aspects of RLHF, so I wonder how I could add more annotators that represent relevant subpopulations, so that the annotator pool is as representative as possible, but without making the annotators behave like stereotypes.

I think the "without...stereotypes" part is a general research question requiring some experimentation (mostly prompting). Note this may be API-model dependent and task-dependent, and ultimately there's no formal guarantee.

Regarding how one can run these prompting experiments with our codebase, you can add your custom prompts here, and change the value to this input argument to point to the folder of prompts and configurations.

Closing this issue now, but feel free to reopen if you have further questions.