You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Based on this thread. We should discuss whether we want to allow people to create custom evaluation procedures (e.g., “30 times 3-fold Crossvalidation”) and how we want to support this. Possible concerns include spotty availability (might not support the entire tasktype x procedure matrix) and splintering experimental results. But there are of course also obvious benefits :)
@PGijsbers
I’m supporting custom eval procedures, but we should keep comparability.
i thhink
allow custom procedures only as a variant of a known base (e.g., base=10‑fold CV, params: folds/repeats/seed), and whitelist task types that can use it. That keeps leaderboards anchored on the base while still allowing flexibility. Happy to help draft a concrete schema/API if this direction makes sense.
Activity
Jayant-kernel commentedon Mar 15, 2026
@PGijsbers
I’m supporting custom eval procedures, but we should keep comparability.
i thhink
allow custom procedures only as a variant of a known base (e.g., base=10‑fold CV, params: folds/repeats/seed), and whitelist task types that can use it. That keeps leaderboards anchored on the base while still allowing flexibility. Happy to help draft a concrete schema/API if this direction makes sense.