promptdojo_promptdojo_promptdojo_promptdojo_promptdojo_promptdojo_promptdojo_promptdojo_promptdojo_promptdojo_promptdojo_

Train/inference skew — step 7 of 7

Checkpoint

One last thing before we move on. pass this to mark the lesson done, or skip and keep moving. hop to the next when you're ready.

Build defense #3: diff serving-time feature distributions against training. For each feature (alphabetical), compute the serving batch's mean, the relative shift abs(serve - train) / train, and print f"{feat:<8} train={t:<6} serve={s:.2f} shift={shift:.0%}" — appending " <- INVESTIGATE" when the shift exceeds 50%. One feature here has quietly quadrupled; the table should make it impossible to miss. (This comparison is chapter 47's drift monitoring being born.)

full-screen editor opens — close anytime to keep reading.