promptdojo_

Overfitting and regularization you can see — step 5 of 7

This tuning loop picked k by the best TEST score — it just spent the test set on a decision, so the 85% it reports is quietly inflated by selection. Knobs get chosen on the validation scores; the test set is opened once, after choosing. Fix the selection line and accept the lower, uninflated number.

The break is on line 5 — but read the whole snippet first.

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