Base64-encoded serialized model artifact.
Wall-clock training time in seconds.
Per-feature importance/contribution.
Fitted preprocessing parameters (means/stds, vocabularies, bin edges, fill
values) learned during /train. Travels with the model and is replayed at
inference — never re-fit (plan §6.2).
Optionalholdout_Honest metrics on the locked holdout, scored exactly once (plan §8.2).
Train + validation metrics.
Number of training rows the model was fit on.
POST /trainresponse. Carries the serialized model plus the fitted preprocessing parameters (the anti-skew payload) and the deterministic metrics used to grade and rank the run.