Results

The latest benchmark numbers committed to the repo.

The full table lives at results/benchmark_results.md and the underlying CSV at results/benchmark_results.csv.

How to read the table

• qini — point estimate on the held-out test fold.
• qini_ci_lower / qini_ci_upper — 95% BCa bootstrap interval (1000 resamples by default).
• auuc_normalized — AUUC divided by perfect-ranking AUUC, in roughly \([-1, 1]\).
• uplift_at_k — realised uplift in the top-k targeted population.
• overlap_ess_ratio — effective sample size after IPW reweighting, divided by N. Closer to 1 = better treatment/control overlap.

Per-dataset breakdowns live in results/<dataset>_results.md.

Reproducing

See Reproducing. Same --seeds give bit-identical numbers; the dataset SHA is logged to MLflow so you can be sure you trained on the same bytes.