Skip to content

Grinsztajn standings

External standings: the Grinsztajn suite

The Grinsztajn et al. tabular benchmark at the paper-medium protocol: 55 OpenML tasks, three seeds, campaign knobs for every library (decision 68). Best variant per library, average rank across tasks, lower is better.

Grinsztajn mean rank by library

library mean rank outright wins
bonsai 1.44 36
lgbm 2.51 5
xgb 2.84 6
catboost 3.22 8

Per-suite mean rank:

library cat_clf cat_reg num_clf num_reg
bonsai 1.29 1.46 1.67 1.30
lgbm 2.29 2.77 2.47 2.45
xgb 3.57 2.54 2.87 2.75
catboost 2.86 3.23 3.00 3.50

Sensitivity: XGBoost's campaign mapping sets min_child_weight=20 (hessian-weighted, the knob-translation bracket recorded in decision 68); replacing its rows with the min_child_weight=1 run gives the other end of the bracket:

library mean rank outright wins
bonsai 1.78 18
xgb 2.18 23
lgbm 2.69 6
catboost 3.35 8

Reproduce: pip install bonsai-gbt[bench], then python -m bonsai.bench.grinsztajn out.jsonl to run the suite (hours; datasets fetch from OpenML), then --report on the same file to render the standings from the jsonl.

Source: grinsztajn-2026-07.jsonl, grinsztajn-2026-07-xgb-mcw1.jsonl. As-run; evidence narrative in benchmarks/grinsztajn-2026-07.md, ruling in decision 68.