Test set submission with pass@n

#35
by caelum19 - opened

Hello!

Is it possible to submit to the test set with pass@n, say pass@3? I think it is meaningful information. For the validation set I am able to get a pass@3 score by comparing with the answers. How private are the test set answers?

For context I am wondering about the scores in the OWL paper and the OAgents paper which separately list their pass@n scores, would they just have been using the validation set? (they do not describe in the papers as far as I could tell), and I am trying to push a small model to its limits, which requires a high temperature in order to avoid loops (even with freq penalty, it gets in semantic loops), where pass@3 is quite a bit harder and seems to be noteworth/meaningful

caelum19 changed discussion title from Test set submission with pass@k to Test set submission with pass@n

Perhaps I should leave this on the leaderboard repo but I guess "How private are the test set answers" is an appropriate question for this repo

Hi! Any paper which reported a pass@n score only submitted on the validation set. The answers of the test set are not public and we only allow pass@1 (strict accuracy)

clefourrier changed discussion status to closed

Sign up or log in to comment