A county assessor parcel desk is training a GPU XGBoost valuation pilot and must decide which knobs count as model parameters needing deliberate tuning. What should they treat as parameters to tune?

A county assessor parcel desk training a GPU XGBoost valuation pilot must treat learning rate, tree depth, and related hyperparameters as the parameters that need deliberate tuning. Those settings control how the boosted trees fit parcel features—how aggressively weights update, how deep splits grow, and how capacity interacts with regularization—so they directly shape bias–variance behavior on assessor data. Tuning them against a held-out validation metric (or cross-validation) is the associate-level practice that turns a pilot into an evidence-based model rather than a one-shot fit on training parcels.

Treating only the Jupyter notebook color theme as a tunable parameter fails by concept: UI cosmetics do not change the hypothesis class or the fit of parcel features, so they cannot improve valuation error. Treating only the street address of the assessor's office fails because facility location metadata is unrelated to model capacity or learning dynamics. Treating only the public-dashboard font size fails for the same reason—presentation styling is not a model parameter and does not govern how XGBoost learns from assessed values or property attributes.

Exam caveat: practitioners sometimes blur learned weights with hyperparameters chosen around training. On this exam, the intended distinction is between settings that control fit behavior—learning rate, depth, and peers—versus irrelevant operational or UI knobs; do not waste search on cosmetics.

Operational check: before a valuation pilot gate review, confirm a small documented hyperparameter search space (learning rate, max depth, and related controls), a fixed validation protocol on parcel holdouts, and that UI or building metadata are excluded from the tuning checklist.