Practise the standard verbs for detecting and responding to ML feature drift.
0 / 5 completed
1 / 5
Fill in: 'We ___ each production feature's distribution against its training-time baseline daily, so a quiet upstream change is caught before it degrades every prediction downstream.'
We 'monitor a distribution' — the standard collocation for ongoing observation of a feature's statistical properties. The other options aren't idiomatic here.
2 / 5
Fill in: 'Skipping drift checks after an upstream schema change can ___ a model quietly scoring on a feature whose meaning shifted weeks ago without anyone noticing.'
We say no drift check will 'leave' a model scoring on a shifted feature — the standard, natural collocation for the resulting risk. The other options aren't idiomatic here.
3 / 5
Fill in: 'We ___ a drifted feature clearly in the monitoring dashboard the moment its statistical distance from the baseline crosses a set threshold, rather than waiting for accuracy to visibly drop.'
We 'flag a feature' — the standard, simple collocation for marking a detected anomaly clearly. The other options are less idiomatic here.
4 / 5
Fill in: 'We ___ the current feature distribution against last month's snapshot rather than only against the original training baseline, so a slow gradual drift doesn't hide behind an outdated comparison.'
We 'compare' distributions — the standard, simple collocation for contrasting two measured statistical snapshots. The other options are less idiomatic here.
5 / 5
Fill in: 'We ___ the model on fresh data the moment sustained drift is confirmed, rather than letting it keep serving predictions built on an assumption that's no longer true.'
We 'retrain a model' — the standard, established collocation for updating a model once its inputs have meaningfully changed. The other options aren't the recognised term here.