hard20 pts · first answer countsmonitoringdrift

Inputs stable but accuracy falls

Feature distribution monitors show no drift, but newly labelled data shows the model's accuracy falling. What is the most likely explanation?

A.Concept drift: P(y|x) changed, which input monitors cannot seeB.Covariate shift in the input featuresC.A tokenizer mismatch between train and serveD.The stored model weights decayed over time

← Filtered vector search returns too few Retraining on your own recommendations →