50 A hyperparameter search in plain terms

What people usually ask

Optimize the hyperparameters of my model.

What comes back

A grid over every parameter, scored on the same rows the final number is quoted from, with nothing saved when the kernel dies.

Ask this instead

I have gran.csv: 48 granulation runs, inputs liquid_pct, impeller_rpm, time_min, target d50_um, and settings.csv with columns name, value, one row per hyperparameter.
1. Load settings.csv with dtype=str. Print each name, its value and type(value). Cast each one on purpose and print the type again, so a size written as "10,10" never becomes the number 1010.
2. An Optuna study: sampler TPESampler(seed=0), storage="sqlite:///study.db", study_name="gran_rf", load_if_exists=True, direction="minimize". Objective: mean RMSE of a RandomForestRegressor over an inner KFold(5) on the training rows only. Search n_estimators 50 to 500, max_depth 2 to 12, min_samples_leaf 1 to 8. Budget: 60 trials.
3. Print each trial's parameters and value as it completes. If the run stops, rerunning the cell resumes from the database.
4. Report the best parameters beside the scikit-learn defaults, with both RMSEs.
5. The score that gets quoted comes from an outer KFold(5) the search never saw: run the search inside each outer fold and report the outer RMSE mean and spread.
Do not touch the outer test rows during the search. Do not drop rows.

Attach: The run table and the settings file, both as CSV.

Why it works: The budget and the database make the search finite and resumable, and the outer loop keeps the quoted score apart from the rows the search tuned on.