Closed riedgar-ms closed 10 months ago
@Harsha-Nori @Nking92 ?
I'm ok with this for now, but in the future I would prefer if we just put everything into 1 file (all_questions.json
). The JSON data for each question should have split and subject properties. That way we will always have only 1 file we look for, and we filter based on the split property if we want a specific split.
The way this is done now, there's an implicit dependency on having a filename formatted like mmlu_{subject}_{split}.json
. In my opinion, subject and split are both data properties and should coincide with the question data itself, not the filename.
It seems like the MMLU testing code already has some assumptions about the filenames, so we would need to change that behavior in addition to this format script.
Make it possible to run
promptbase
on the entire MMLU dataset via anall
subject. This takes a simple "disk space is cheap" approach.