Model comparison (/p/<project>/bakeoff)

Compare trained and baseline models on evaluation datasets in one run. Present only when the backend mounts its bake-off router — absent, not disabled, otherwise.

1. Pick evaluation datasets (export test splits, or external frozen sets grouped by the backend's own grouping).
2. Pick models (finished training runs, baseline models, or a custom reference), each showing per-dataset facts and a train/test overlap warning when relevant.
3. Pick a scoring profile.
4. Confirm and run — the job polls with progress and per-stage failures.

Results

A model × dataset matrix with every served tied winner bolded, plus ranked per-dataset results and a per-class table (an uncovered class reads "not covered", not a blank). A pre-v2 result shows a compatibility note instead of a broken render. Nothing here computes a metric, mapping, rank or winner client-side — it's all served.

Bake-off — datasets, models and the results matrix