bench-picker

Run 5 benchmarks instead of 57.

— models — tasks last refresh: —

1 · Pick a family

2 · Choose subset size

5 of —

3 · Or target R²

R² ≥ 0.90

Submodular MI greedy (Krause/Singh/Guestrin 2008). Held-out R² via leave-one-model-out ridge regression. All scores fetched live from public leaderboards.

?

Pick a family on the left, choose k, and hit Compute. We'll greedy-select the k tasks that maximize information gain over the rest, then verify by predicting every held-out task with ridge regression.