One split instead of k folds: fit every param map once, and keep the best.

Latu.ML.CrossValidator with num_folds: 1 is not this — a cross-validation scores every param map k times and averages. This scores each one exactly once, against a single held-out slice, and is what you reach for when a fit is expensive enough that k of them is not worth the variance reduction.

tvs =
  ML.train_validation_split(
    estimator: lr,
    param_maps: ML.param_grid(lr, reg_param: [0.1, 0.01]),
    evaluator: Evaluation.binary_classification_evaluator(),
    train_ratio: 0.75
  )

{:ok, model} = ML.fit(tvs, training)

The split is cut the same way a fold is — rand(seed) and a range — so train_ratio: 0.75 means rows whose draw is under 0.75, not exactly three quarters of them. Pass seed: for a split that is the same twice.