Field note · August 24, 2023
Cutting a lesson down — Validation that matches deployment
Cross-validation is not one technique. Choosing the wrong variant is how models get shipped with an offline score that has no relationship to what happens next.
A note from writing Validation that matches deployment, which took three passes to get to something short.
Cross-validation is not one technique. Choosing the wrong variant is how models get shipped with an offline score that has no relationship to what happens next.
The hard part of writing this was cutting it. The first version covered every case; the useful version covers the case you hit on a Tuesday and names the rest in a sentence. Completeness is a property of reference material, not of teaching material, and confusing the two produces something nobody finishes.
It sits in the Modelling, honestly module. The exercise is a thinking one rather than a query one — some of this material is about deciding what to compute, and that does not need a keyboard.
Ordering matters here more than in most courses. It follows Frame the problem, then beat a stupid baseline and leads into Features, and explaining what the model did, and reading it out of sequence mostly works but costs you the setup.
Free means free, and it also means we can rewrite it whenever it is wrong. No edition, no errata PDF, nothing to repurchase. The whole course is 36 lessons and the fixes land the day we find them.