Which tuned hyperparameters stop being correct once you refit on the full training set?

## The gap nobody looks at Selection happens at one training-set size and deployment happens at another. With 5-fold cross-validation every candidate was judged by models trained on 80% of the rows; the artifact you ship trains on 100%. That is a 25% increase in training data between the measurement and the thing measured. Most hyperparameters do not care. A few care a lot, and copying those across is one of the quieter ways a refit ends up worse than the fold models it replaced. ## Which settings move with n, and in which direction **Number of boosting rounds (gradient-boosted trees).** Each round adds capacity. With more training data, the point at which additional rounds start fitting noise arrives later, so the optimal count generally **rises** with n. A count tuned on 80% of the rows therefore tends to underfit the full refit. A widely used heuristic is to scale the tuned count by `k/(k-1)` — a factor of 1.25 at k=5, 1.11 at k=10 — on the argument that the refit sees that much more data. It is a rough correction, not a law: with a small learning rate the loss curve is flat near its minimum and the exact count barely matters, while with an aggressive learning rate it matters a great deal. **Neighbour count k in a nearest-neighbour model.** For consistency the neighbourhood must grow in absolute size while shrinking as a share of the sample — `k` increases with n while `k/n` goes to zero. Practically, a k tuned at n=8,000 is modestly too small at n=10,000: the same k now spans a physically smaller region, giving a slightly noisier, lower-bias estimate than the tuning suggested. Scaling with the square root of n is the usual rule of thumb, which would nudge k up by roughly 12% for that 25% increase in rows. **Regularization strength.** A penalty exists to control variance. Variance falls as n grows, so the optimal penalty typically **weakens** with more data — the fold-tuned value is usually a little too strong for the refit. The effect is mild for a 25% jump and material when the jump is large. **Anything written as an absolute count.** A minimum of 20 samples per leaf is a constraint of 20/6,400 of the data during a fold and 20/8,000 after the refit: the same number is a *weaker* brake at larger n, so the refit tree grows deeper than the tuning implied. Early-stopping windows measured in rounds, minimum samples per split, and minimum counts for a category to survive encoding all share this property. ## Which settings travel unchanged Settings that describe the *shape* of the hypothesis rather than how much evidence supports it: - the learning rate and the loss function, - the distance metric or kernel choice, - the encoding strategy for categorical features, - the number of trees in a random forest — more trees reduce prediction variance and do not overfit, they only cost time, - anything already expressed as a fraction: a minimum leaf as a share of n, a row-subsample or column-subsample rate, a validation share. That last bullet is the design lesson. **A setting expressed as a proportion rescales itself; a setting expressed as a count does not.** Where a learner offers both forms, prefer the proportional one and the whole problem disappears. ## Three ways to handle the size-sensitive ones 1. **Rescale by the size ratio.** Multiply round counts by `k/(k-1)`, nudge k for neighbours by the square-root rule, ease the penalty slightly. Cheap, approximate, and much better than doing nothing. 2. **Re-derive on the full set.** Carve a small internal validation slice out of the full training data and read the size-sensitive value off it directly, keeping every other setting at its cross-validated value. Costs one extra fit and gives the value at the correct sample size. 3. **Tune only proportional forms.** Restate the search space so nothing is size-dependent in the first place. This is the version that survives contact with a team, because it needs no one to remember the correction. ## The tell in an interview A candidate who says "the hyperparameters transfer, that is the whole point of tuning" has never watched a refit get worse. A candidate who says "I audit the setting list before the refit and split it into size-sensitive and size-neutral" has. The direction matters too: more data supports **more** capacity and **less** penalty. Getting that backwards — claiming a bigger training set needs stronger regularization — is the misconception this question is built to catch. ## Keeping it in proportion For a 25% increase in rows the corrections are usually small, and on a flat loss surface they may be invisible. The point is not that the refit will fail without them; it is that you know which numbers are assumptions about sample size and can say so. On a bigger jump — say tuning on a 10% subsample for speed and refitting on everything — the same corrections stop being cosmetic and start deciding whether the model works.