medium10 pts · first answer countsci-cdmodel-evaluation

Automatic gate before promotion

A weekly retrain promotes models automatically. Which gate is most useful before promotion?

A.Promote if training loss went downB.Compare the candidate with the current model on a fixed held-out set and key slices, and block it if it is worse beyond a toleranceC.Promote if the model file is newerD.Promote if training finished without errors

← Deploying by alias, not version Training on spot GPUs →