Databricks model catalog

Choose a model by its behavior, then copy one of its models: entries into config/training.yml. This catalog covers all 16 regression and 18 classification choices in the project wizard, including voting and stacking.

Keep the surrounding version: 1, defaults: and your existing data, target, validation and layout settings from the training reference. Each example below replaces the models: mapping; do not concatenate the tuned and untuned mappings. In a competition, give candidates distinct names. In a multi-target project, retain each branch's target and feature settings.

The untuned example fixes the estimator configuration. The tuned example uses a deliberately small explicit grid to demonstrate valid YAML; it is not the full automatic search space. Validation/CV comes from defaults, not from a tuning.cv field. Read search strategies and ensembles for all strategy fields, nested CV, budgets, member parameters and wizard questions.

Reading defaults and search spaces

The parameter tables are generated from the same Core hyperparameter registry used by the wizard's build_model_spaces.py. They show key supported controls, not every parameter accepted by an upstream estimator. Core catalog default is the registered control value, not a promise about an omitted YAML field. The adjacent column is generated from the calculator's actual constructor settings. Library/resolver means the calculator delegates that default to the installed estimator or its composite resolver. For example, Logistic Regression registers max_iter=100 in the catalog but an omitted YAML value runs with Core's max_iter=1000. Random Forest has similar explicit Core overrides. Supply a parameter explicitly when reproducing a catalog setting matters. Core also injects its deterministic seed into compatible estimators when no seed is given.

Table literals use JSON notation (null, true, false); the explicit constructor column preserves Python source (None, True, False). Use YAML null, true, false in configuration. A table's default does not imply that all combinations of its controls are valid; solver/penalty pairs and estimator capabilities are still checked.

Grid and halving-grid use the compact grid column. Random, halving-random and Optuna use the richer finite-choice column when the wizard supplies its automatic space. The wizard writes these axes into the generated configuration. A missing or empty YAML search_space also asks the runtime to derive the Core space for the selected model, respecting its fixed parameters and ensemble members. For halving, the default resource is n_samples; a resource parameter cannot also appear in its search space. The strategy guide explains minimum/maximum resources and how to limit the full Cartesian grid. Large automatic grids, especially ensembles with several members, can exceed max_candidates before fitting. Use a bounded explicit space such as the examples below, or a random strategy with an explicit trial budget.

Weights, optional packages and execution boundaries

weight_column is a training workflow setting, not a constructor parameter. The column is removed from model features; weights must be finite, nonnegative, and have a positive sum in every fitted fold. Neighbor models do not accept row weights. Ensembles require every selected base model and the final estimator to support them. Metrics and threshold selection remain unweighted. An ensemble's weights controls member averaging and is a separate concept.

Every card's Spark/serving note describes the current admission policy, not proof that a particular saved artifact will pass it. Training these estimators uses local pandas/Polars data within the configured row/byte bounds; choosing Spark inference does not turn fitting into Spark ML training. Spark worker and pinned serving admission inspect a complete pandas-fitted artifact, exact model classes, preprocessing, schemas, runtime and saved state. An otherwise eligible model can still be rejected by its preprocessing or custom scoring rules. See Spark execution and serving rollout.

Optional XGBoost/LightGBM packages must be installed in training and inference environments. Installing a package does not extend the admission allowlist. Local fit/predict examples do not establish Spark, REST, or ai_query acceptance.

ridge_regression

Ridge regression. Task: regression. A linear baseline with L2 shrinkage, useful when numeric predictors are correlated or there are many modest signals. Scale numeric features before comparing alpha values.

Sample weights: Yes; configure the workflow weight_column.

Optional package: None beyond the standard Core scikit-learn runtime.

Spark workers / serving: Not admitted by the current Spark-worker/pinned-serving model allowlist; use local inference.

Parameters and defaults

Parameter	Core catalog default	Explicit Core constructor default	Meaning
alpha	1.0	1.0	Regularization strength; must be a positive float.
solver	"auto"	'auto'	Solver to use in the computational routines.
fit_intercept	true	Library/resolver	Whether to calculate the intercept for this model.
random_state	42	Library/resolver	Seed used by the stochastic sag/saga solvers.

Default search spaces

Axis	Grid / halving_grid	Random / halving_random / Optuna
alpha	[0.01, 0.1, 1.0, 10.0, 100.0]	[0.01, 0.1, 1.0, 10.0, 100.0]
fit_intercept	[true, false]	[true, false]
max_iter	[2000]	[1000, 2000, 5000]
solver	["auto", "saga"]	["auto", "svd", "cholesky", "lsqr", "sparse_cg", "sag", "saga"]

Halving resource: n_samples (the wizard default). This model has no top-level n_estimators resource. Use n_samples in the wizard; another estimator parameter needs an explicit compatible configuration.

Minimal training entries

Untuned:

models:
  candidate:
    model:
      type: ridge_regression
      params: {}

Tuned (small explicit grid):

models:
  candidate:
    model:
      type: ridge_regression
      params: {}
    tuning:
      strategy: grid
      metric: r2
      max_candidates: 4
      search_space:
        alpha: [0.01, 0.1]

Common mistake: Assuming the same alpha has the same effect before and after feature scaling.

linear_regression

Ordinary least squares. Task: regression. A transparent baseline when a numeric target changes approximately linearly with the features. Coefficients are useful for understanding direction, provided the features have meaningful units.

Sample weights: Yes; configure the workflow weight_column.

Optional package: None beyond the standard Core scikit-learn runtime.

Spark workers / serving: Eligible exact model family, subject to complete artifact admission.

Parameters and defaults

Parameter	Core catalog default	Explicit Core constructor default	Meaning
fit_intercept	true	True	Whether to calculate the intercept for this model.
copy_X	Not a catalog control	True	Core constructor setting.
n_jobs	Not a catalog control	-1	Core constructor setting.

Default search spaces

Axis	Grid / halving_grid	Random / halving_random / Optuna
fit_intercept	[true, false]	[true, false]

Halving resource: n_samples (the wizard default). This model has no top-level n_estimators resource. Use n_samples in the wizard; another estimator parameter needs an explicit compatible configuration.

Minimal training entries

Untuned:

models:
  candidate:
    model:
      type: linear_regression
      params: {}

Tuned (small explicit grid):

models:
  candidate:
    model:
      type: linear_regression
      params: {}
    tuning:
      strategy: grid
      metric: r2
      max_candidates: 4
      search_space:
        fit_intercept: [true, false]

Common mistake: Treating correlated coefficients as independent causal effects; remove redundant features or compare Ridge.

adaboost_regressor

AdaBoost regression. Task: regression. Sequentially reweights the learning problem to focus later learners on difficult examples. A compact boosting baseline when simple weak learners capture useful signal.

Sample weights: Yes; configure the workflow weight_column.

Optional package: None beyond the standard Core scikit-learn runtime.

Spark workers / serving: Not admitted by the current Spark-worker/pinned-serving model allowlist; use local inference.

Parameters and defaults

Parameter	Core catalog default	Explicit Core constructor default	Meaning
n_estimators	50	50	The maximum number of estimators at which boosting is terminated.
learning_rate	1.0	1.0	Weight applied to each classifier at each boosting iteration.
random_state	42	Library/resolver	Seed passed to the base estimator's randomness, if it has any.

Default search spaces

Axis	Grid / halving_grid	Random / halving_random / Optuna
learning_rate	[0.1, 0.5, 1.0]	[0.01, 0.1, 1.0, 1.5]
n_estimators	[50, 100, 200]	[50, 100, 200, 500]

Halving resource: n_samples (the wizard default). This model also exposes n_estimators; selecting it requires explicit max_resources and removing that axis from the search space.

Minimal training entries

Untuned:

models:
  candidate:
    model:
      type: adaboost_regressor
      params: {}

Tuned (small explicit grid):

models:
  candidate:
    model:
      type: adaboost_regressor
      params: {}
    tuning:
      strategy: grid
      metric: r2
      max_candidates: 4
      search_space:
        n_estimators: [50, 100]

Common mistake: Allowing mislabeled rows or severe target outliers to dominate the repeatedly reweighted fit.

decision_tree_regressor

Decision tree regression. Task: regression. A readable nonlinear baseline that splits feature space into regions. Depth and leaf constraints control how finely it partitions the data.

Sample weights: Yes; configure the workflow weight_column.

Optional package: None beyond the standard Core scikit-learn runtime.

Spark workers / serving: Eligible exact model family, subject to complete artifact admission.

Parameters and defaults

Parameter	Core catalog default	Explicit Core constructor default	Meaning
max_depth	null	None	The maximum depth of the tree.
min_samples_split	2	2	The minimum number of samples required to split an internal node.
min_samples_leaf	1	Library/resolver	The minimum number of samples required to be at a leaf node.
random_state	42	Library/resolver	Seed for split tie-breaking and feature sub-sampling.
criterion	"squared_error"	'squared_error'	The function to measure the quality of a split.

Default search spaces

Axis	Grid / halving_grid	Random / halving_random / Optuna
criterion	["squared_error", "friedman_mse"]	["squared_error", "friedman_mse", "absolute_error"]
max_depth	[5, 10, 20]	[null, 3, 5, 10, 20]
min_samples_leaf	[]	[1, 2, 4]
min_samples_split	[2, 5, 10]	[2, 5, 10]

Halving resource: n_samples (the wizard default). This model has no top-level n_estimators resource. Use n_samples in the wizard; another estimator parameter needs an explicit compatible configuration.

Minimal training entries

Untuned:

models:
  candidate:
    model:
      type: decision_tree_regressor
      params: {}

Tuned (small explicit grid):

models:
  candidate:
    model:
      type: decision_tree_regressor
      params: {}
    tuning:
      strategy: grid
      metric: r2
      max_candidates: 4
      search_space:
        max_depth: [5, 10]

Common mistake: Leaving a deep tree unconstrained on a small noisy dataset and treating the training score as generalization.

elasticnet_regression

Elastic Net regression. Task: regression. Combines L1 feature selection and L2 shrinkage. Useful when correlated groups of features make pure Lasso selection unstable; alpha controls total shrinkage and l1_ratio its mix.

Sample weights: Yes; configure the workflow weight_column.

Optional package: None beyond the standard Core scikit-learn runtime.

Spark workers / serving: Not admitted by the current Spark-worker/pinned-serving model allowlist; use local inference.

Parameters and defaults

Parameter	Core catalog default	Explicit Core constructor default	Meaning
alpha	1.0	1.0	Constant that multiplies the penalty terms.
l1_ratio	0.5	0.5	The ElasticNet mixing parameter (0 <= l1_ratio <= 1).
selection	"cyclic"	'cyclic'	If set to 'random', a random coefficient is updated every iteration.
random_state	42	Library/resolver	Seed for the coefficient update order when selection='random'.

Default search spaces

Axis	Grid / halving_grid	Random / halving_random / Optuna
alpha	[0.01, 0.1, 1.0]	[0.001, 0.01, 0.1, 1.0, 10.0]
l1_ratio	[0.3, 0.5, 0.7]	[0.1, 0.5, 0.7, 0.9]

Halving resource: n_samples (the wizard default). This model has no top-level n_estimators resource. Use n_samples in the wizard; another estimator parameter needs an explicit compatible configuration.

Minimal training entries

Untuned:

models:
  candidate:
    model:
      type: elasticnet_regression
      params: {}

Tuned (small explicit grid):

models:
  candidate:
    model:
      type: elasticnet_regression
      params: {}
    tuning:
      strategy: grid
      metric: r2
      max_candidates: 4
      search_space:
        alpha: [0.01, 0.1]

Common mistake: Tuning l1_ratio without tuning alpha or scaling the input features.

extra_trees_regressor

Extremely randomized trees regression. Task: regression. Averages trees with extra split randomization, providing a useful tree-ensemble comparison to Random Forest. Its default bootstrap is false.

Sample weights: Yes; configure the workflow weight_column.

Optional package: None beyond the standard Core scikit-learn runtime.

Spark workers / serving: Eligible exact model family, subject to complete artifact admission.

Parameters and defaults

Parameter	Core catalog default	Explicit Core constructor default	Meaning
n_estimators	100	100	Number of trees in the forest.
max_depth	null	None	Maximum depth of the tree. None = unlimited.
min_samples_split	2	2	Minimum samples required to split an internal node.
min_samples_leaf	1	1	Minimum samples required at a leaf node.
bootstrap	false	False	Whether bootstrap samples are used (False = use full dataset).
random_state	42	Library/resolver	Seed for bootstrap sampling and random split selection.
criterion	"squared_error"	'squared_error'	Function to measure the quality of a split.
n_jobs	Not a catalog control	-1	Core constructor setting.

Default search spaces

Axis	Grid / halving_grid	Random / halving_random / Optuna
bootstrap	[]	[false, true]
criterion	["squared_error", "friedman_mse"]	["squared_error", "absolute_error", "friedman_mse"]
max_depth	[5, 10, 20]	[null, 5, 10, 20, 50]
min_samples_leaf	[]	[1, 2, 4]
min_samples_split	[2, 5]	[2, 5, 10]
n_estimators	[100, 200, 500]	[50, 100, 200, 500]

Halving resource: n_samples (the wizard default). This model also exposes n_estimators; selecting it requires explicit max_resources and removing that axis from the search space.

Minimal training entries

Untuned:

models:
  candidate:
    model:
      type: extra_trees_regressor
      params: {}

Tuned (small explicit grid):

models:
  candidate:
    model:
      type: extra_trees_regressor
      params: {}
    tuning:
      strategy: grid
      metric: r2
      max_candidates: 4
      search_space:
        n_estimators: [100, 200]

Common mistake: Assuming bootstrap sampling is always active, or that more trees will fix excessive leaf-level overfitting.

gradient_boosting_regressor

Gradient boosting regression. Task: regression. Builds small trees sequentially to correct earlier errors. Useful for tabular nonlinear relationships; learning_rate and n_estimators jointly set the fitting budget.

Sample weights: Yes; configure the workflow weight_column.

Optional package: None beyond the standard Core scikit-learn runtime.

Spark workers / serving: Not admitted by the current Spark-worker/pinned-serving model allowlist; use local inference.

Parameters and defaults

Parameter	Core catalog default	Explicit Core constructor default	Meaning
n_estimators	100	100	The number of boosting stages to perform.
learning_rate	0.1	0.1	Shrinks the contribution of each tree by learning_rate.
max_depth	3	3	Maximum depth of the individual regression estimators.
subsample	1.0	Library/resolver	The fraction of samples to be used for fitting the individual base learners.
min_samples_split	2	Library/resolver	Minimum samples required to split an internal node.
min_samples_leaf	1	Library/resolver	Minimum samples required at a leaf node.
random_state	42	Library/resolver	Seed for the row subsampling used when subsample < 1.0.

Default search spaces

Axis	Grid / halving_grid	Random / halving_random / Optuna
learning_rate	[0.05, 0.1, 0.2]	[0.01, 0.05, 0.1, 0.2]
max_depth	[3, 5]	[3, 5, 7, 9]
min_samples_leaf	[1, 5]	[1, 5, 10, 20]
n_estimators	[100, 200, 500]	[50, 100, 200, 500]
subsample	[0.8, 1.0]	[0.6, 0.8, 1.0]

Halving resource: n_samples (the wizard default). This model also exposes n_estimators; selecting it requires explicit max_resources and removing that axis from the search space.

Minimal training entries

Untuned:

models:
  candidate:
    model:
      type: gradient_boosting_regressor
      params: {}

Tuned (small explicit grid):

models:
  candidate:
    model:
      type: gradient_boosting_regressor
      params: {}
    tuning:
      strategy: grid
      metric: r2
      max_candidates: 4
      search_space:
        n_estimators: [100, 200]

Common mistake: Increasing learning_rate and tree count together without validation, which can make overfitting much worse.

hist_gradient_boosting_regressor

Histogram gradient boosting regression. Task: regression. Uses feature bins to speed up boosted-tree fitting on larger tabular datasets. max_iter is the boosting-round count; leaves and regularization control capacity.

Sample weights: Yes; configure the workflow weight_column.

Optional package: None beyond the standard Core scikit-learn runtime.

Spark workers / serving: Not admitted by the current Spark-worker/pinned-serving model allowlist; use local inference.

Parameters and defaults

Parameter	Core catalog default	Explicit Core constructor default	Meaning
max_iter	100	100	Maximum number of iterations (boosting rounds).
learning_rate	0.1	0.1	Shrinks the contribution of each tree.
max_leaf_nodes	31	31	Maximum number of leaves per tree.
max_depth	null	None	Maximum depth per tree. None = unlimited (max_leaf_nodes is the effective limit).
min_samples_leaf	20	20	Minimum samples at a leaf node.
l2_regularization	0.0	0.0	L2 penalty applied to the leaves' values.
max_bins	255	255	Maximum number of bins for feature discretisation. Higher = more precise but slower.
random_state	42	Library/resolver	Seed for the histogram-based gradient boosting randomness.

Default search spaces

Axis	Grid / halving_grid	Random / halving_random / Optuna
l2_regularization	[]	[0.0, 0.01, 0.1, 1.0]
learning_rate	[0.05, 0.1, 0.2]	[0.01, 0.05, 0.1, 0.2]
max_depth	[]	[null, 3, 5, 10]
max_iter	[100, 200, 500]	[50, 100, 200, 500]
max_leaf_nodes	[31, 63]	[15, 31, 63, 127]
min_samples_leaf	[20, 50]	[10, 20, 50, 100]

Halving resource: n_samples (the wizard default). This model has no top-level n_estimators resource. Use n_samples in the wizard; another estimator parameter needs an explicit compatible configuration.

Minimal training entries

Untuned:

models:
  candidate:
    model:
      type: hist_gradient_boosting_regressor
      params: {}

Tuned (small explicit grid):

models:
  candidate:
    model:
      type: hist_gradient_boosting_regressor
      params: {}
    tuning:
      strategy: grid
      metric: r2
      max_candidates: 4
      search_space:
        max_iter: [100, 200]

Common mistake: Using n_estimators as a constructor field or halving resource; this family uses max_iter and the wizard default resource is n_samples.

k_neighbors_regressor

Nearest-neighbor regression. Task: regression. Predicts by averaging nearby training targets. A useful local-pattern baseline when distance between scaled feature vectors has a clear meaning; prediction retains dependence on the training set.

Sample weights: No; neighbor-distance weights is a different setting.

Optional package: None beyond the standard Core scikit-learn runtime.

Spark workers / serving: Not admitted by the current Spark-worker/pinned-serving model allowlist; use local inference.

Parameters and defaults

Parameter	Core catalog default	Explicit Core constructor default	Meaning
n_neighbors	5	5	Number of neighbors to use.
weights	"uniform"	'uniform'	Weight function used in prediction.
algorithm	"auto"	'auto'	Algorithm used to compute the nearest neighbors.
n_jobs	Not a catalog control	-1	Core constructor setting.

Default search spaces

Axis	Grid / halving_grid	Random / halving_random / Optuna
algorithm	["auto"]	["auto", "ball_tree", "kd_tree", "brute"]
n_neighbors	[3, 5, 9]	[3, 5, 7, 9, 11]
weights	["uniform", "distance"]	["uniform", "distance"]

Halving resource: n_samples (the wizard default). This model has no top-level n_estimators resource. Use n_samples in the wizard; another estimator parameter needs an explicit compatible configuration.

Minimal training entries

Untuned:

models:
  candidate:
    model:
      type: k_neighbors_regressor
      params: {}

Tuned (small explicit grid):

models:
  candidate:
    model:
      type: k_neighbors_regressor
      params: {}
    tuning:
      strategy: grid
      metric: r2
      max_candidates: 4
      search_space:
        n_neighbors: [3, 5]

Common mistake: Confusing distance-based neighbor weights with sample_weight; this estimator cannot fit explicit row weights.

lasso_regression

Lasso regression. Task: regression. L1 regularization can shrink some coefficients to zero, giving a compact linear model when only a subset of features carries signal. Scale inputs and check convergence.

Sample weights: Yes; configure the workflow weight_column.

Optional package: None beyond the standard Core scikit-learn runtime.

Spark workers / serving: Not admitted by the current Spark-worker/pinned-serving model allowlist; use local inference.

Parameters and defaults

Parameter	Core catalog default	Explicit Core constructor default	Meaning
alpha	1.0	1.0	Constant that multiplies the L1 term.
selection	"cyclic"	'cyclic'	If set to 'random', a random coefficient is updated every iteration.
random_state	42	Library/resolver	Seed for the coefficient update order when selection='random'.

Default search spaces

Axis	Grid / halving_grid	Random / halving_random / Optuna
alpha	[0.001, 0.01, 0.1, 1.0, 10.0]	[0.001, 0.01, 0.1, 1.0, 10.0]

Halving resource: n_samples (the wizard default). This model has no top-level n_estimators resource. Use n_samples in the wizard; another estimator parameter needs an explicit compatible configuration.

Minimal training entries

Untuned:

models:
  candidate:
    model:
      type: lasso_regression
      params: {}

Tuned (small explicit grid):

models:
  candidate:
    model:
      type: lasso_regression
      params: {}
    tuning:
      strategy: grid
      metric: r2
      max_candidates: 4
      search_space:
        alpha: [0.001, 0.01]

Common mistake: Reading a zero coefficient as proof that the feature has no signal when correlated substitutes are present.

lgbm_regressor

LightGBM regression. Task: regression. An optional histogram-based boosted-tree model. num_leaves controls leaf-wise model capacity; min_child_samples helps keep small leaves from overfitting.

Sample weights: Yes; configure the workflow weight_column.

Optional package: lightgbm.

Spark workers / serving: Not admitted by the current Spark-worker/pinned-serving model allowlist; use local inference.

Parameters and defaults

Parameter	Core catalog default	Explicit Core constructor default	Meaning
n_estimators	100	100	Number of boosting rounds.
num_leaves	31	31	Maximum number of leaves per tree. Controls model complexity; increase for more accuracy.
learning_rate	0.1	0.1	Shrinks the contribution of each tree.
max_depth	-1	-1	Maximum tree depth. -1 = unlimited (controlled by num_leaves).
min_child_samples	20	20	Minimum samples required in a leaf. Regularises against overfitting.
subsample	1.0	1.0	Fraction of training rows sampled each iteration for GBDT/DART. GOSS uses gradient sampling instead.
colsample_bytree	1.0	1.0	Fraction of features sampled per tree.
reg_alpha	0.0	0.0	L1 regularization term on leaf weights.
reg_lambda	0.0	0.0	L2 regularization term on leaf weights.
boosting_type	"gbdt"	'gbdt'	GBDT is standard gradient boosting. DART adds dropout. GOSS samples by gradient magnitude.
random_state	42	Library/resolver	Seed for feature/row subsampling (bagging, colsample, DART).
n_jobs	Not a catalog control	-1	Core constructor setting.
subsample_freq	Not a catalog control	None	Core constructor setting.
verbose	Not a catalog control	-1	Core constructor setting.
verbosity	Not a catalog control	-1	Core constructor setting.

Default search spaces

Axis	Grid / halving_grid	Random / halving_random / Optuna
boosting_type	[]	["gbdt", "dart", "goss"]
colsample_bytree	[]	[0.6, 0.8, 1.0]
learning_rate	[0.05, 0.1, 0.2]	[0.01, 0.05, 0.1, 0.2]
max_depth	[-1, 5, 10]	[-1, 5, 10, 20]
min_child_samples	[]	[5, 10, 20, 50]
n_estimators	[100, 200, 500]	[100, 200, 500, 1000]
num_leaves	[31, 63, 127]	[15, 31, 63, 127, 255]
reg_alpha	[]	[0.0, 0.01, 0.1, 1.0]
reg_lambda	[]	[0.0, 0.01, 0.1, 1.0]
subsample	[0.8, 1.0]	[0.6, 0.8, 1.0]

Halving resource: n_samples (the wizard default). This model also exposes n_estimators; selecting it requires explicit max_resources and removing that axis from the search space.

Minimal training entries

Untuned:

models:
  candidate:
    model:
      type: lgbm_regressor
      params: {}

Tuned (small explicit grid):

models:
  candidate:
    model:
      type: lgbm_regressor
      params: {}
    tuning:
      strategy: grid
      metric: r2
      max_candidates: 4
      search_space:
        n_estimators: [100, 200]

Common mistake: Setting a large num_leaves on a small dataset, or assuming library availability alone grants Spark or serving admission.

random_forest_regressor

Random forest regression. Task: regression. Averages many randomized trees to capture nonlinear interactions with less variance than a single tree. It usually does not need numeric feature scaling.

Sample weights: Yes; configure the workflow weight_column.

Optional package: None beyond the standard Core scikit-learn runtime.

Spark workers / serving: Eligible exact model family, subject to complete artifact admission.

Parameters and defaults

Parameter	Core catalog default	Explicit Core constructor default	Meaning
n_estimators	100	50	The number of trees in the forest.
max_depth	null	10	The maximum depth of the tree. If None, nodes are expanded until all leaves are pure.
min_samples_split	2	5	The minimum number of samples required to split an internal node.
min_samples_leaf	1	2	The minimum number of samples required to be at a leaf node.
bootstrap	true	Library/resolver	Whether bootstrap samples are used when building trees.
random_state	42	Library/resolver	Seed for bootstrap sampling and tree building. Same data + same seed = identical forest.
n_jobs	Not a catalog control	-1	Core constructor setting.

Default search spaces

Axis	Grid / halving_grid	Random / halving_random / Optuna
bootstrap	[]	[true, false]
criterion	["squared_error", "friedman_mse"]	["squared_error", "friedman_mse", "absolute_error"]
max_depth	[5, 10, 20]	[null, 5, 10, 20, 30, 50]
min_samples_leaf	[1, 4]	[1, 2, 4, 8]
min_samples_split	[2, 10]	[2, 5, 10, 20]
n_estimators	[100, 200, 500]	[50, 100, 200, 500]

Halving resource: n_samples (the wizard default). This model also exposes n_estimators; selecting it requires explicit max_resources and removing that axis from the search space.

Minimal training entries

Untuned:

models:
  candidate:
    model:
      type: random_forest_regressor
      params: {}

Tuned (small explicit grid):

models:
  candidate:
    model:
      type: random_forest_regressor
      params: {}
    tuning:
      strategy: grid
      metric: r2
      max_candidates: 4
      search_space:
        n_estimators: [100, 200]

Common mistake: Assuming an omitted parameter means the upstream library default: Core starts at 50 trees, depth 10, split size 5 and leaf size 2.

stacking_regressor

Stacking ensemble regression. Task: regression. Learns a final estimator from out-of-fold base predictions. Its internal cv prevents the final learner from training on overoptimistic in-sample predictions; passthrough also includes original features.

Sample weights: Yes for the default configuration; every selected child and final estimator must support it.

Optional package: None beyond the standard Core scikit-learn runtime; add xgboost or lightgbm when selected as a member.

Spark workers / serving: Not admitted by the current Spark-worker/pinned-serving model allowlist; use local inference.

Parameters and defaults

Parameter	Core catalog default	Explicit Core constructor default	Meaning
base_estimators	["random_forest", "gradient_boosting", "ridge"]	Library/resolver	Base learners combined by the ensemble.
final_estimator	"ridge"	Library/resolver	Meta-learner trained on the base models' out-of-fold predictions.
cv	5	5	Cross-validation folds used to generate out-of-fold predictions for the final estimator. Keep small (e.g. 3) when also running an outer hyperparameter search to limit nested-CV cost.
passthrough	false	Library/resolver	When on, the final estimator sees the original input features alongside the base models' predictions instead of the predictions alone. Can help when the base learners miss signal in the raw data.
n_jobs	1	Library/resolver	Number of base models fit in parallel. 1 = sequential, -1 = use all CPU cores. Speeds up fitting when you have several independent base learners.

Default search spaces

These axes use the catalog's default members shown above. The wizard expands your selected members instead; it keeps composition, voting and internal CV fixed. Calibrated members add estimator__ after their name.

Axis	Grid / halving_grid	Random / halving_random / Optuna
final_estimator__alpha	[0.01, 0.1, 1.0, 10.0, 100.0]	[0.01, 0.1, 1.0, 10.0, 100.0]
final_estimator__fit_intercept	[true, false]	[true, false]
final_estimator__max_iter	[2000]	[1000, 2000, 5000]
final_estimator__solver	["auto", "saga"]	["auto", "svd", "cholesky", "lsqr", "sparse_cg", "sag", "saga"]
gradient_boosting__learning_rate	[0.05, 0.1, 0.2]	[0.01, 0.05, 0.1, 0.2]
gradient_boosting__max_depth	[3, 5]	[3, 5, 7, 9]
gradient_boosting__min_samples_leaf	[1, 5]	[1, 5, 10, 20]
gradient_boosting__n_estimators	[100, 200, 500]	[50, 100, 200, 500]
gradient_boosting__subsample	[0.8, 1.0]	[0.6, 0.8, 1.0]
random_forest__bootstrap	[]	[true, false]
random_forest__criterion	["squared_error", "friedman_mse"]	["squared_error", "friedman_mse", "absolute_error"]
random_forest__max_depth	[5, 10, 20]	[null, 5, 10, 20, 30, 50]
random_forest__min_samples_leaf	[1, 4]	[1, 2, 4, 8]
random_forest__min_samples_split	[2, 10]	[2, 5, 10, 20]
random_forest__n_estimators	[100, 200, 500]	[50, 100, 200, 500]
ridge__alpha	[0.01, 0.1, 1.0, 10.0, 100.0]	[0.01, 0.1, 1.0, 10.0, 100.0]
ridge__fit_intercept	[true, false]	[true, false]
ridge__max_iter	[2000]	[1000, 2000, 5000]
ridge__solver	["auto", "saga"]	["auto", "svd", "cholesky", "lsqr", "sparse_cg", "sag", "saga"]

Halving resource: n_samples (the wizard default). This model has no top-level n_estimators resource. Use n_samples in the wizard; another estimator parameter needs an explicit compatible configuration.

Minimal training entries

Untuned:

models:
  candidate:
    model:
      type: stacking_regressor
      params: {"base_estimators": ["ridge", "decision_tree"], "final_estimator": "ridge", "cv": 3}

Tuned (small explicit grid):

models:
  candidate:
    model:
      type: stacking_regressor
      params: {"base_estimators": ["ridge", "decision_tree"], "final_estimator": "ridge", "cv": 3}
    tuning:
      strategy: grid
      metric: r2
      max_candidates: 4
      search_space:
        ridge__alpha: [0.01, 0.1]

Common mistake: Confusing internal cv with the external model-selection folds, or multiplying both by a large search budget without checking the resulting fit count.

svr

Support vector regression. Task: regression. Kernel regression for moderate-sized datasets with smooth nonlinear relationships. Scale features and compare the kernel, C and gamma together; large datasets can make fitting expensive.

Sample weights: Yes; configure the workflow weight_column.

Optional package: None beyond the standard Core scikit-learn runtime.

Spark workers / serving: Not admitted by the current Spark-worker/pinned-serving model allowlist; use local inference.

Parameters and defaults

Parameter	Core catalog default	Explicit Core constructor default	Meaning
C	1.0	1.0	Regularization parameter. The strength of the regularization is inversely proportional to C.
kernel	"rbf"	'rbf'	Specifies the kernel type to be used in the algorithm.
gamma	"scale"	'scale'	Kernel coefficient for 'rbf', 'poly' and 'sigmoid'.

Default search spaces

Axis	Grid / halving_grid	Random / halving_random / Optuna
C	[0.1, 1.0, 10.0]	[0.1, 1.0, 10.0, 100.0]
gamma	["scale", "auto"]	["scale", "auto"]
kernel	["linear", "rbf"]	["linear", "rbf", "poly", "sigmoid"]

Halving resource: n_samples (the wizard default). This model has no top-level n_estimators resource. Use n_samples in the wizard; another estimator parameter needs an explicit compatible configuration.

Minimal training entries

Untuned:

models:
  candidate:
    model:
      type: svr
      params: {}

Tuned (small explicit grid):

models:
  candidate:
    model:
      type: svr
      params: {}
    tuning:
      strategy: grid
      metric: r2
      max_candidates: 4
      search_space:
        C: [0.1, 1.0]

Common mistake: Applying an RBF kernel to features with wildly different units, or assuming its training cost scales like a linear model.

voting_regressor

Voting ensemble regression. Task: regression. Combines independently fitted base learners. Regression averages predictions; classification can average probabilities (soft) or use a majority label vote (hard). weights follow the order of base_estimators.

Sample weights: Yes for the default configuration; every selected child and final estimator must support it.

Optional package: None beyond the standard Core scikit-learn runtime; add xgboost or lightgbm when selected as a member.

Spark workers / serving: Not admitted by the current Spark-worker/pinned-serving model allowlist; use local inference.

Parameters and defaults

Parameter	Core catalog default	Explicit Core constructor default	Meaning
base_estimators	["linear_regression", "random_forest", "gradient_boosting"]	Library/resolver	Fixed base learners combined by the ensemble; enable base model tuning to search their parameters.
n_jobs	1	Library/resolver	Number of base models fit in parallel. 1 = sequential, -1 = use all CPU cores. Speeds up fitting when you have several independent base learners.

Default search spaces

These axes use the catalog's default members shown above. The wizard expands your selected members instead; it keeps composition, voting and internal CV fixed. Calibrated members add estimator__ after their name.

Axis	Grid / halving_grid	Random / halving_random / Optuna
gradient_boosting__learning_rate	[0.05, 0.1, 0.2]	[0.01, 0.05, 0.1, 0.2]
gradient_boosting__max_depth	[3, 5]	[3, 5, 7, 9]
gradient_boosting__min_samples_leaf	[1, 5]	[1, 5, 10, 20]
gradient_boosting__n_estimators	[100, 200, 500]	[50, 100, 200, 500]
gradient_boosting__subsample	[0.8, 1.0]	[0.6, 0.8, 1.0]
linear_regression__fit_intercept	[true, false]	[true, false]
random_forest__bootstrap	[]	[true, false]
random_forest__criterion	["squared_error", "friedman_mse"]	["squared_error", "friedman_mse", "absolute_error"]
random_forest__max_depth	[5, 10, 20]	[null, 5, 10, 20, 30, 50]
random_forest__min_samples_leaf	[1, 4]	[1, 2, 4, 8]
random_forest__min_samples_split	[2, 10]	[2, 5, 10, 20]
random_forest__n_estimators	[100, 200, 500]	[50, 100, 200, 500]

Halving resource: n_samples (the wizard default). This model has no top-level n_estimators resource. Use n_samples in the wizard; another estimator parameter needs an explicit compatible configuration.

Minimal training entries

Untuned:

models:
  candidate:
    model:
      type: voting_regressor
      params: {"base_estimators": ["ridge", "decision_tree"]}

Tuned (small explicit grid):

models:
  candidate:
    model:
      type: voting_regressor
      params: {"base_estimators": ["ridge", "decision_tree"]}
    tuning:
      strategy: grid
      metric: r2
      max_candidates: 4
      search_space:
        ridge__alpha: [0.01, 0.1]

Common mistake: Using full registry IDs as member names: use ridge or random_forest, for example, rather than ridge_regression or random_forest_regressor. Voting regression averages numeric predictions; member weights must follow the member order.

xgboost_regressor

XGBoost regression. Task: regression. An optional gradient-boosted tree model with row/column sampling and regularization controls. Useful for tabular interactions when the matching xgboost package is available in every runtime.

Sample weights: Yes; configure the workflow weight_column.

Optional package: xgboost.

Spark workers / serving: Eligible exact model family, subject to complete artifact admission. Only reviewed CPU gbtree regressors with supported built-in objectives qualify.

Parameters and defaults

Parameter	Core catalog default	Explicit Core constructor default	Meaning
n_estimators	100	100	Number of gradient boosted trees.
max_depth	6	6	Maximum tree depth for base learners.
learning_rate	0.3	0.3	Boosting learning rate (eta).
subsample	1.0	0.8	Subsample ratio of the training instances.
colsample_bytree	1.0	0.8	Subsample ratio of columns when constructing each tree.
min_child_weight	1	Library/resolver	Minimum sum of instance weights in a child. Higher values regularize against overfitting.
gamma	0.0	Library/resolver	Minimum loss reduction required to make a split. Higher = more conservative tree growth.
reg_alpha	0.0	Library/resolver	L1 regularization term on leaf weights.
reg_lambda	1.0	Library/resolver	L2 regularization term on leaf weights. Default sklearn XGBoost is 1.
random_state	42	Library/resolver	Seed for row/column subsampling when subsample or colsample_bytree < 1.0.
n_jobs	Not a catalog control	-1	Core constructor setting.

Default search spaces

Axis	Grid / halving_grid	Random / halving_random / Optuna
colsample_bytree	[]	[0.6, 0.8, 1.0]
learning_rate	[0.05, 0.1, 0.2]	[0.01, 0.05, 0.1, 0.3]
max_depth	[3, 5, 7]	[3, 5, 7, 9]
min_child_weight	[1, 3, 5]	[1, 3, 5, 7]
n_estimators	[100, 200, 500]	[100, 200, 500, 1000]
reg_alpha	[]	[0.0, 0.01, 0.1, 1.0]
reg_lambda	[0.1, 1.0, 5.0]	[0.1, 1.0, 5.0, 10.0]
subsample	[0.8, 1.0]	[0.6, 0.8, 1.0]

Halving resource: n_samples (the wizard default). This model also exposes n_estimators; selecting it requires explicit max_resources and removing that axis from the search space.

Minimal training entries

Untuned:

models:
  candidate:
    model:
      type: xgboost_regressor
      params: {}

Tuned (small explicit grid):

models:
  candidate:
    model:
      type: xgboost_regressor
      params: {}
    tuning:
      strategy: grid
      metric: r2
      max_candidates: 4
      search_space:
        n_estimators: [100, 200]

Common mistake: Installing xgboost only on the training driver, or assuming every classifier/booster configuration is admitted for Spark workers.

logistic_regression

Logistic regression. Task: classification. A probabilistic linear classification baseline for binary or multiclass targets. Scale numeric features; C is inverse regularization strength, so smaller C means stronger shrinkage.

Sample weights: Yes; configure the workflow weight_column.

Optional package: None beyond the standard Core scikit-learn runtime.

Spark workers / serving: Eligible exact model family, subject to complete artifact admission.

Parameters and defaults

Parameter	Core catalog default	Explicit Core constructor default	Meaning
C	1.0	Library/resolver	Smaller values specify stronger regularization.
penalty	"l2"	Library/resolver	Norm used in the penalization.
solver	"lbfgs"	'lbfgs'	Algorithm to use in the optimization problem.
max_iter	100	1000	Maximum number of iterations taken for the solvers to converge.
l1_ratio	null	Library/resolver	The ElasticNet mixing parameter, with 0 <= l1_ratio <= 1. Only used if penalty='elasticnet'.
random_state	42	Library/resolver	Seed used by stochastic solvers (saga, liblinear shuffling).

Default search spaces

Axis	Grid / halving_grid	Random / halving_random / Optuna
C	[0.01, 0.1, 1.0, 10.0]	[0.001, 0.01, 0.1, 1.0, 10.0, 100.0]
max_iter	[500, 2000]	[100, 200, 500, 1000, 2000, 5000]
penalty	["l1", "l2"]	["l1", "l2"]
solver	["saga"]	["saga"]

Halving resource: n_samples (the wizard default). This model has no top-level n_estimators resource. Use n_samples in the wizard; another estimator parameter needs an explicit compatible configuration.

Minimal training entries

Untuned:

models:
  candidate:
    model:
      type: logistic_regression
      params: {}

Tuned (small explicit grid):

models:
  candidate:
    model:
      type: logistic_regression
      params: {}
    tuning:
      strategy: grid
      metric: accuracy
      max_candidates: 4
      search_space:
        C: [0.01, 0.1]

Common mistake: Choosing an incompatible solver and penalty pair. The default generated grid switches to saga so its L1 and L2 candidates are compatible.

adaboost_classifier

AdaBoost classification. Task: classification. Sequentially reweights the learning problem to focus later learners on difficult examples. A compact boosting baseline when simple weak learners capture useful signal.

Sample weights: Yes; configure the workflow weight_column.

Optional package: None beyond the standard Core scikit-learn runtime.

Spark workers / serving: Not admitted by the current Spark-worker/pinned-serving model allowlist; use local inference.

Parameters and defaults

Parameter	Core catalog default	Explicit Core constructor default	Meaning
n_estimators	50	50	The maximum number of estimators at which boosting is terminated.
learning_rate	1.0	1.0	Weight applied to each classifier at each boosting iteration.
random_state	42	Library/resolver	Seed passed to the base estimator's randomness, if it has any.

Default search spaces

Axis	Grid / halving_grid	Random / halving_random / Optuna
learning_rate	[0.1, 0.5, 1.0]	[0.01, 0.1, 1.0, 1.5]
n_estimators	[50, 100, 200]	[50, 100, 200, 500]

Halving resource: n_samples (the wizard default). This model also exposes n_estimators; selecting it requires explicit max_resources and removing that axis from the search space.

Minimal training entries

Untuned:

models:
  candidate:
    model:
      type: adaboost_classifier
      params: {}

Tuned (small explicit grid):

models:
  candidate:
    model:
      type: adaboost_classifier
      params: {}
    tuning:
      strategy: grid
      metric: accuracy
      max_candidates: 4
      search_space:
        n_estimators: [50, 100]

Common mistake: Allowing mislabeled rows or severe target outliers to dominate the repeatedly reweighted fit.

bernoulli_nb

Bernoulli naive Bayes. Task: classification. A classifier for binary occurrence features, such as whether a token is present. binarize defines the threshold applied to numeric inputs; null means inputs are already binary.

Sample weights: Yes; configure the workflow weight_column.

Optional package: None beyond the standard Core scikit-learn runtime.

Spark workers / serving: Not admitted by the current Spark-worker/pinned-serving model allowlist; use local inference.

Parameters and defaults

Parameter	Core catalog default	Explicit Core constructor default	Meaning
alpha	1.0	1.0	Additive (Laplace/Lidstone) smoothing parameter.
binarize	0.0	0.0	Threshold for binarizing features. Features above this value are set to 1, others to 0. Set to None to skip binarization (assume features are already binary).
fit_prior	true	True	Whether to learn class prior probabilities.

Default search spaces

Axis	Grid / halving_grid	Random / halving_random / Optuna
alpha	[0.1, 0.5, 1.0, 2.0]	[0.01, 0.1, 0.5, 1.0, 2.0, 5.0]
binarize	[0.0, 0.5]	[0.0, 0.5, 1.0]
fit_prior	[]	[true, false]

Halving resource: n_samples (the wizard default). This model has no top-level n_estimators resource. Use n_samples in the wizard; another estimator parameter needs an explicit compatible configuration.

Minimal training entries

Untuned:

models:
  candidate:
    model:
      type: bernoulli_nb
      params: {}

Tuned (small explicit grid):

models:
  candidate:
    model:
      type: bernoulli_nb
      params: {}
    tuning:
      strategy: grid
      metric: accuracy
      max_candidates: 4
      search_space:
        alpha: [0.1, 0.5]

Common mistake: Using count magnitude as a signal while binarization discards it; consider MultinomialNB for counts.

calibrated_classifier

Calibrated classifier. Task: classification. Wraps a base classifier with an internal cross-validated probability calibrator. Useful when probability quality matters in addition to class ranking; sigmoid is a compact starting point.

Sample weights: Yes for the default configuration; every selected child and final estimator must support it.

Optional package: None beyond the standard Core scikit-learn runtime.

Spark workers / serving: Not admitted by the current Spark-worker/pinned-serving model allowlist; use local inference.

Parameters and defaults

Parameter	Core catalog default	Explicit Core constructor default	Meaning
base_estimator	"logistic_regression"	Library/resolver	Classifier whose probabilities are calibrated. Uses sensible defaults for the chosen estimator. During tuning, each selected classifier is a search candidate.
method	"sigmoid"	'sigmoid'	How to map raw scores to calibrated probabilities. Sigmoid (Platt) suits small datasets; isotonic is non-parametric and needs more data.
cv	5	5	Number of cross-validation folds used to fit the calibrator.
random_state	42	Library/resolver	Seed for stochastic base estimators. Does not change calibration CV splits or deterministic base estimators.
estimator	Not a catalog control	LogisticRegression(max_iter=1000)	Core constructor setting.

Default search spaces

Axis	Grid / halving_grid	Random / halving_random / Optuna
cv	[3, 5]	[3, 5, 10]
method	["sigmoid", "isotonic"]	["sigmoid", "isotonic"]

Halving resource: n_samples (the wizard default). This model has no top-level n_estimators resource. Use n_samples in the wizard; another estimator parameter needs an explicit compatible configuration.

Minimal training entries

Untuned:

models:
  candidate:
    model:
      type: calibrated_classifier
      params: {}

Tuned (small explicit grid):

models:
  candidate:
    model:
      type: calibrated_classifier
      params: {}
    tuning:
      strategy: grid
      metric: accuracy
      max_candidates: 4
      search_space:
        method: ["sigmoid", "isotonic"]

Common mistake: Using isotonic calibration with very little data, or forgetting that calibration cv adds fits inside an outer search.

decision_tree_classifier

Decision tree classification. Task: classification. A readable nonlinear baseline that splits feature space into regions. Depth and leaf constraints control how finely it partitions the data.

Sample weights: Yes; configure the workflow weight_column.

Optional package: None beyond the standard Core scikit-learn runtime.

Spark workers / serving: Eligible exact model family, subject to complete artifact admission.

Parameters and defaults

Parameter	Core catalog default	Explicit Core constructor default	Meaning
max_depth	null	None	The maximum depth of the tree.
min_samples_split	2	2	The minimum number of samples required to split an internal node.
min_samples_leaf	1	Library/resolver	The minimum number of samples required to be at a leaf node.
random_state	42	Library/resolver	Seed for split tie-breaking and feature sub-sampling.
criterion	"gini"	'gini'	The function to measure the quality of a split.

Default search spaces

Axis	Grid / halving_grid	Random / halving_random / Optuna
criterion	["gini", "entropy"]	["gini", "entropy"]
max_depth	[5, 10, 20]	[null, 3, 5, 10, 20]
min_samples_leaf	[]	[1, 2, 4]
min_samples_split	[2, 5, 10]	[2, 5, 10]

Halving resource: n_samples (the wizard default). This model has no top-level n_estimators resource. Use n_samples in the wizard; another estimator parameter needs an explicit compatible configuration.

Minimal training entries

Untuned:

models:
  candidate:
    model:
      type: decision_tree_classifier
      params: {}

Tuned (small explicit grid):

models:
  candidate:
    model:
      type: decision_tree_classifier
      params: {}
    tuning:
      strategy: grid
      metric: accuracy
      max_candidates: 4
      search_space:
        max_depth: [5, 10]

Common mistake: Leaving a deep tree unconstrained on a small noisy dataset and treating the training score as generalization.

extra_trees_classifier

Extremely randomized trees classification. Task: classification. Averages trees with extra split randomization, providing a useful tree-ensemble comparison to Random Forest. Its default bootstrap is false.

Sample weights: Yes; configure the workflow weight_column.

Optional package: None beyond the standard Core scikit-learn runtime.

Spark workers / serving: Eligible exact model family, subject to complete artifact admission.

Parameters and defaults

Parameter	Core catalog default	Explicit Core constructor default	Meaning
n_estimators	100	100	Number of trees in the forest.
max_depth	null	None	Maximum depth of the tree. None = unlimited.
min_samples_split	2	2	Minimum samples required to split an internal node.
min_samples_leaf	1	1	Minimum samples required at a leaf node.
bootstrap	false	False	Whether bootstrap samples are used (False = use full dataset).
random_state	42	Library/resolver	Seed for bootstrap sampling and random split selection.
criterion	"gini"	'gini'	Function to measure the quality of a split.
n_jobs	Not a catalog control	-1	Core constructor setting.

Default search spaces

Axis	Grid / halving_grid	Random / halving_random / Optuna
bootstrap	[]	[false, true]
criterion	["gini", "entropy"]	["gini", "entropy", "log_loss"]
max_depth	[5, 10, 20]	[null, 5, 10, 20, 50]
min_samples_leaf	[]	[1, 2, 4]
min_samples_split	[2, 5]	[2, 5, 10]
n_estimators	[100, 200, 500]	[50, 100, 200, 500]

Halving resource: n_samples (the wizard default). This model also exposes n_estimators; selecting it requires explicit max_resources and removing that axis from the search space.

Minimal training entries

Untuned:

models:
  candidate:
    model:
      type: extra_trees_classifier
      params: {}

Tuned (small explicit grid):

models:
  candidate:
    model:
      type: extra_trees_classifier
      params: {}
    tuning:
      strategy: grid
      metric: accuracy
      max_candidates: 4
      search_space:
        n_estimators: [100, 200]

Common mistake: Assuming bootstrap sampling is always active, or that more trees will fix excessive leaf-level overfitting.

gaussian_nb

Gaussian naive Bayes. Task: classification. A fast classification baseline for continuous features, modeling a separate Gaussian distribution per feature and class. Its conditional-independence assumption can still give useful classification.

Sample weights: Yes; configure the workflow weight_column.

Optional package: None beyond the standard Core scikit-learn runtime.

Spark workers / serving: Not admitted by the current Spark-worker/pinned-serving model allowlist; use local inference.

Parameters and defaults

Parameter	Core catalog default	Explicit Core constructor default	Meaning
var_smoothing	1e-09	1e-09	Portion of the largest variance of all features that is added to variances for calculation stability.

Default search spaces

Axis	Grid / halving_grid	Random / halving_random / Optuna
var_smoothing	[1e-09, 1e-08, 1e-07, 1e-06]	[1e-09, 1e-08, 1e-07, 1e-06]

Halving resource: n_samples (the wizard default). This model has no top-level n_estimators resource. Use n_samples in the wizard; another estimator parameter needs an explicit compatible configuration.

Minimal training entries

Untuned:

models:
  candidate:
    model:
      type: gaussian_nb
      params: {}

Tuned (small explicit grid):

models:
  candidate:
    model:
      type: gaussian_nb
      params: {}
    tuning:
      strategy: grid
      metric: accuracy
      max_candidates: 4
      search_space:
        var_smoothing: [1e-09, 1e-08]

Common mistake: Interpreting confident probabilities as calibrated when highly correlated features repeat the same evidence.

gradient_boosting_classifier

Gradient boosting classification. Task: classification. Builds small trees sequentially to correct earlier errors. Useful for tabular nonlinear relationships; learning_rate and n_estimators jointly set the fitting budget.

Sample weights: Yes; configure the workflow weight_column.

Optional package: None beyond the standard Core scikit-learn runtime.

Spark workers / serving: Not admitted by the current Spark-worker/pinned-serving model allowlist; use local inference.

Parameters and defaults

Parameter	Core catalog default	Explicit Core constructor default	Meaning
n_estimators	100	100	The number of boosting stages to perform.
learning_rate	0.1	0.1	Shrinks the contribution of each tree by learning_rate.
max_depth	3	3	Maximum depth of the individual regression estimators.
subsample	1.0	Library/resolver	The fraction of samples to be used for fitting the individual base learners.
min_samples_split	2	Library/resolver	Minimum samples required to split an internal node.
min_samples_leaf	1	Library/resolver	Minimum samples required at a leaf node.
random_state	42	Library/resolver	Seed for the row subsampling used when subsample < 1.0.

Default search spaces

Axis	Grid / halving_grid	Random / halving_random / Optuna
learning_rate	[0.05, 0.1, 0.2]	[0.01, 0.05, 0.1, 0.2]
max_depth	[3, 5]	[3, 5, 7, 9]
min_samples_leaf	[1, 5]	[1, 5, 10, 20]
n_estimators	[100, 200, 500]	[50, 100, 200, 500]
subsample	[0.8, 1.0]	[0.6, 0.8, 1.0]

Halving resource: n_samples (the wizard default). This model also exposes n_estimators; selecting it requires explicit max_resources and removing that axis from the search space.

Minimal training entries

Untuned:

models:
  candidate:
    model:
      type: gradient_boosting_classifier
      params: {}

Tuned (small explicit grid):

models:
  candidate:
    model:
      type: gradient_boosting_classifier
      params: {}
    tuning:
      strategy: grid
      metric: accuracy
      max_candidates: 4
      search_space:
        n_estimators: [100, 200]

Common mistake: Increasing learning_rate and tree count together without validation, which can make overfitting much worse.

hist_gradient_boosting_classifier

Histogram gradient boosting classification. Task: classification. Uses feature bins to speed up boosted-tree fitting on larger tabular datasets. max_iter is the boosting-round count; leaves and regularization control capacity.

Sample weights: Yes; configure the workflow weight_column.

Optional package: None beyond the standard Core scikit-learn runtime.

Spark workers / serving: Not admitted by the current Spark-worker/pinned-serving model allowlist; use local inference.

Parameters and defaults

Parameter	Core catalog default	Explicit Core constructor default	Meaning
max_iter	100	100	Maximum number of iterations (boosting rounds).
learning_rate	0.1	0.1	Shrinks the contribution of each tree.
max_leaf_nodes	31	31	Maximum number of leaves per tree.
max_depth	null	None	Maximum depth per tree. None = unlimited (max_leaf_nodes is the effective limit).
min_samples_leaf	20	20	Minimum samples at a leaf node.
l2_regularization	0.0	0.0	L2 penalty applied to the leaves' values.
max_bins	255	255	Maximum number of bins for feature discretisation. Higher = more precise but slower.
random_state	42	Library/resolver	Seed for the histogram-based gradient boosting randomness.

Default search spaces

Axis	Grid / halving_grid	Random / halving_random / Optuna
l2_regularization	[]	[0.0, 0.01, 0.1, 1.0]
learning_rate	[0.05, 0.1, 0.2]	[0.01, 0.05, 0.1, 0.2]
max_depth	[]	[null, 3, 5, 10]
max_iter	[100, 200, 500]	[50, 100, 200, 500]
max_leaf_nodes	[31, 63]	[15, 31, 63, 127]
min_samples_leaf	[20, 50]	[10, 20, 50, 100]

Halving resource: n_samples (the wizard default). This model has no top-level n_estimators resource. Use n_samples in the wizard; another estimator parameter needs an explicit compatible configuration.

Minimal training entries

Untuned:

models:
  candidate:
    model:
      type: hist_gradient_boosting_classifier
      params: {}

Tuned (small explicit grid):

models:
  candidate:
    model:
      type: hist_gradient_boosting_classifier
      params: {}
    tuning:
      strategy: grid
      metric: accuracy
      max_candidates: 4
      search_space:
        max_iter: [100, 200]

Common mistake: Using n_estimators as a constructor field or halving resource; this family uses max_iter and the wizard default resource is n_samples.

k_neighbors_classifier

Nearest-neighbor classifier. Task: classification. Classifies from nearby labeled rows. Useful as a distance-based baseline with appropriately scaled features and enough neighbors from every class.

Sample weights: No; neighbor-distance weights is a different setting.

Optional package: None beyond the standard Core scikit-learn runtime.

Spark workers / serving: Not admitted by the current Spark-worker/pinned-serving model allowlist; use local inference.

Parameters and defaults

Parameter	Core catalog default	Explicit Core constructor default	Meaning
n_neighbors	5	5	Number of neighbors to use.
weights	"uniform"	'uniform'	Weight function used in prediction.
algorithm	"auto"	'auto'	Algorithm used to compute the nearest neighbors.
n_jobs	Not a catalog control	-1	Core constructor setting.

Default search spaces

Axis	Grid / halving_grid	Random / halving_random / Optuna
algorithm	["auto"]	["auto", "ball_tree", "kd_tree", "brute"]
n_neighbors	[3, 5, 9]	[3, 5, 7, 9, 11]
weights	["uniform", "distance"]	["uniform", "distance"]

Halving resource: n_samples (the wizard default). This model has no top-level n_estimators resource. Use n_samples in the wizard; another estimator parameter needs an explicit compatible configuration.

Minimal training entries

Untuned:

models:
  candidate:
    model:
      type: k_neighbors_classifier
      params: {}

Tuned (small explicit grid):

models:
  candidate:
    model:
      type: k_neighbors_classifier
      params: {}
    tuning:
      strategy: grid
      metric: accuracy
      max_candidates: 4
      search_space:
        n_neighbors: [3, 5]

Common mistake: Setting n_neighbors above the number of rows in the smallest CV training fold, or expecting row-weight support.

lgbm_classifier

LightGBM classification. Task: classification. An optional histogram-based boosted-tree model. num_leaves controls leaf-wise model capacity; min_child_samples helps keep small leaves from overfitting.

Sample weights: Yes; configure the workflow weight_column.

Optional package: lightgbm.

Spark workers / serving: Not admitted by the current Spark-worker/pinned-serving model allowlist; use local inference.

Parameters and defaults

Parameter	Core catalog default	Explicit Core constructor default	Meaning
n_estimators	100	100	Number of boosting rounds.
num_leaves	31	31	Maximum number of leaves per tree. Controls model complexity; increase for more accuracy.
learning_rate	0.1	0.1	Shrinks the contribution of each tree.
max_depth	-1	-1	Maximum tree depth. -1 = unlimited (controlled by num_leaves).
min_child_samples	20	20	Minimum samples required in a leaf. Regularises against overfitting.
subsample	1.0	1.0	Fraction of training rows sampled each iteration for GBDT/DART. GOSS uses gradient sampling instead.
colsample_bytree	1.0	1.0	Fraction of features sampled per tree.
reg_alpha	0.0	0.0	L1 regularization term on leaf weights.
reg_lambda	0.0	0.0	L2 regularization term on leaf weights.
boosting_type	"gbdt"	'gbdt'	GBDT is standard gradient boosting. DART adds dropout. GOSS samples by gradient magnitude.
random_state	42	Library/resolver	Seed for feature/row subsampling (bagging, colsample, DART).
class_weight	null	Library/resolver	Automatically adjusts weights inversely proportional to class frequencies in the training data, useful for imbalanced targets.
n_jobs	Not a catalog control	-1	Core constructor setting.
subsample_freq	Not a catalog control	None	Core constructor setting.
verbose	Not a catalog control	-1	Core constructor setting.
verbosity	Not a catalog control	-1	Core constructor setting.

Default search spaces

Axis	Grid / halving_grid	Random / halving_random / Optuna
boosting_type	[]	["gbdt", "dart", "goss"]
colsample_bytree	[]	[0.6, 0.8, 1.0]
learning_rate	[0.05, 0.1, 0.2]	[0.01, 0.05, 0.1, 0.2]
max_depth	[-1, 5, 10]	[-1, 5, 10, 20]
min_child_samples	[]	[5, 10, 20, 50]
n_estimators	[100, 200, 500]	[100, 200, 500, 1000]
num_leaves	[31, 63, 127]	[15, 31, 63, 127, 255]
reg_alpha	[]	[0.0, 0.01, 0.1, 1.0]
reg_lambda	[]	[0.0, 0.01, 0.1, 1.0]
subsample	[0.8, 1.0]	[0.6, 0.8, 1.0]

Halving resource: n_samples (the wizard default). This model also exposes n_estimators; selecting it requires explicit max_resources and removing that axis from the search space.

Minimal training entries

Untuned:

models:
  candidate:
    model:
      type: lgbm_classifier
      params: {}

Tuned (small explicit grid):

models:
  candidate:
    model:
      type: lgbm_classifier
      params: {}
    tuning:
      strategy: grid
      metric: accuracy
      max_candidates: 4
      search_space:
        n_estimators: [100, 200]

Common mistake: Setting a large num_leaves on a small dataset, or assuming library availability alone grants Spark or serving admission.

multinomial_nb

Multinomial naive Bayes. Task: classification. A fast classifier for nonnegative counts or frequency-like features, such as token counts. alpha smooths rare feature counts and fit_prior controls learned class priors.

Sample weights: Yes; configure the workflow weight_column.

Optional package: None beyond the standard Core scikit-learn runtime.

Spark workers / serving: Not admitted by the current Spark-worker/pinned-serving model allowlist; use local inference.

Parameters and defaults

Parameter	Core catalog default	Explicit Core constructor default	Meaning
alpha	1.0	1.0	Additive (Laplace/Lidstone) smoothing parameter. 0 = no smoothing.
fit_prior	true	True	Whether to learn class prior probabilities. If False, uniform priors are used.

Default search spaces

Axis	Grid / halving_grid	Random / halving_random / Optuna
alpha	[0.1, 0.5, 1.0, 2.0]	[0.01, 0.1, 0.5, 1.0, 2.0, 5.0]
fit_prior	[true, false]	[true, false]

Halving resource: n_samples (the wizard default). This model has no top-level n_estimators resource. Use n_samples in the wizard; another estimator parameter needs an explicit compatible configuration.

Minimal training entries

Untuned:

models:
  candidate:
    model:
      type: multinomial_nb
      params: {}

Tuned (small explicit grid):

models:
  candidate:
    model:
      type: multinomial_nb
      params: {}
    tuning:
      strategy: grid
      metric: accuracy
      max_candidates: 4
      search_space:
        alpha: [0.1, 0.5]

Common mistake: Passing negative values produced by centering or standard scaling; use nonnegative features.

random_forest_classifier

Random forest classification. Task: classification. Averages many randomized trees to capture nonlinear interactions with less variance than a single tree. It usually does not need numeric feature scaling.

Sample weights: Yes; configure the workflow weight_column.

Optional package: None beyond the standard Core scikit-learn runtime.

Spark workers / serving: Eligible exact model family, subject to complete artifact admission.

Parameters and defaults

Parameter	Core catalog default	Explicit Core constructor default	Meaning
n_estimators	100	50	The number of trees in the forest.
max_depth	null	10	The maximum depth of the tree. If None, nodes are expanded until all leaves are pure.
min_samples_split	2	5	The minimum number of samples required to split an internal node.
min_samples_leaf	1	2	The minimum number of samples required to be at a leaf node.
bootstrap	true	Library/resolver	Whether bootstrap samples are used when building trees.
random_state	42	Library/resolver	Seed for bootstrap sampling and tree building. Same data + same seed = identical forest.
criterion	"gini"	Library/resolver	The function to measure the quality of a split.
class_weight	null	Library/resolver	Automatically adjusts weights inversely proportional to class frequencies in the training data, useful for imbalanced targets.
n_jobs	Not a catalog control	-1	Core constructor setting.

Default search spaces

Axis	Grid / halving_grid	Random / halving_random / Optuna
bootstrap	[]	[true, false]
criterion	["gini", "entropy"]	["gini", "entropy", "log_loss"]
max_depth	[5, 10, 20]	[null, 5, 10, 20, 30, 50]
min_samples_leaf	[1, 4]	[1, 2, 4, 8]
min_samples_split	[2, 10]	[2, 5, 10, 20]
n_estimators	[100, 200, 500]	[50, 100, 200, 500]

Halving resource: n_samples (the wizard default). This model also exposes n_estimators; selecting it requires explicit max_resources and removing that axis from the search space.

Minimal training entries

Untuned:

models:
  candidate:
    model:
      type: random_forest_classifier
      params: {}

Tuned (small explicit grid):

models:
  candidate:
    model:
      type: random_forest_classifier
      params: {}
    tuning:
      strategy: grid
      metric: accuracy
      max_candidates: 4
      search_space:
        n_estimators: [100, 200]

Common mistake: Assuming an omitted parameter means the upstream library default: Core starts at 50 trees, depth 10, split size 5 and leaf size 2.

sgd_classifier

Stochastic gradient classifier. Task: classification. Fits a linear decision rule with stochastic optimization, useful for large feature matrices. The Core default loss is log_loss, which exposes probabilities; other losses change the output capabilities.

Sample weights: Yes; configure the workflow weight_column.

Optional package: None beyond the standard Core scikit-learn runtime.

Spark workers / serving: Not admitted by the current Spark-worker/pinned-serving model allowlist; use local inference.

Parameters and defaults

Parameter	Core catalog default	Explicit Core constructor default	Meaning
loss	"log_loss"	'log_loss'	The loss function to be used.
penalty	"l2"	'l2'	The regularization term to use.
alpha	0.0001	0.0001	Constant that multiplies the regularization term. Smaller means weaker regularization.
l1_ratio	0.15	0.15	The ElasticNet mixing parameter, with 0 <= l1_ratio <= 1. Only used if penalty='elasticnet'.
max_iter	1000	1000	The maximum number of passes over the training data (epochs).
random_state	42	Library/resolver	Seed for the SGD shuffle and weight initialization.

Default search spaces

Axis	Grid / halving_grid	Random / halving_random / Optuna
alpha	[0.0001, 0.001]	[1e-05, 0.0001, 0.001, 0.01]
loss	["log_loss", "hinge"]	["log_loss", "hinge", "modified_huber"]
max_iter	[1000]	[500, 1000, 2000]
penalty	["l2"]	["l1", "l2"]

Halving resource: n_samples (the wizard default). This model has no top-level n_estimators resource. Use n_samples in the wizard; another estimator parameter needs an explicit compatible configuration.

Minimal training entries

Untuned:

models:
  candidate:
    model:
      type: sgd_classifier
      params: {}

Tuned (small explicit grid):

models:
  candidate:
    model:
      type: sgd_classifier
      params: {}
    tuning:
      strategy: grid
      metric: accuracy
      max_candidates: 4
      search_space:
        loss: ["log_loss", "hinge"]

Common mistake: Selecting hinge loss while expecting predict_proba or soft voting. Check probability requirements when changing loss.

stacking_classifier

Stacking ensemble classification. Task: classification. Learns a final estimator from out-of-fold base predictions. Its internal cv prevents the final learner from training on overoptimistic in-sample predictions; passthrough also includes original features.

Sample weights: Yes for the default configuration; every selected child and final estimator must support it.

Optional package: None beyond the standard Core scikit-learn runtime; add xgboost or lightgbm when selected as a member.

Spark workers / serving: Not admitted by the current Spark-worker/pinned-serving model allowlist; use local inference.

Parameters and defaults

Parameter	Core catalog default	Explicit Core constructor default	Meaning
base_estimators	["random_forest", "gradient_boosting", "svc"]	Library/resolver	Base learners combined by the ensemble.
final_estimator	"logistic_regression"	Library/resolver	Meta-learner trained on the base models' out-of-fold predictions.
cv	5	5	Cross-validation folds used to generate out-of-fold predictions for the final estimator. Keep small (e.g. 3) when also running an outer hyperparameter search to limit nested-CV cost.
passthrough	false	Library/resolver	When on, the final estimator sees the original input features alongside the base models' predictions instead of the predictions alone. Can help when the base learners miss signal in the raw data.
n_jobs	1	Library/resolver	Number of base models fit in parallel. 1 = sequential, -1 = use all CPU cores. Speeds up fitting when you have several independent base learners.
calibrate_base_models	false	Library/resolver	Wrap each base classifier in CalibratedClassifierCV so its predicted probabilities are well-calibrated. Improves soft voting and stacking when base models give over/under-confident probabilities.
calibration_method	"sigmoid"	Library/resolver	Sigmoid (Platt scaling) is robust on small data; isotonic is more flexible but needs more samples to avoid overfitting.
calibration_cv	3	Library/resolver	Cross-validation folds used to fit each base model's calibrator.

Default search spaces

These axes use the catalog's default members shown above. The wizard expands your selected members instead; it keeps composition, voting and internal CV fixed. Calibrated members add estimator__ after their name.

Axis	Grid / halving_grid	Random / halving_random / Optuna
final_estimator__C	[0.01, 0.1, 1.0, 10.0]	[0.001, 0.01, 0.1, 1.0, 10.0, 100.0]
final_estimator__max_iter	[500, 2000]	[100, 200, 500, 1000, 2000, 5000]
final_estimator__penalty	["l1", "l2"]	["l1", "l2"]
final_estimator__solver	["saga"]	["saga"]
gradient_boosting__learning_rate	[0.05, 0.1, 0.2]	[0.01, 0.05, 0.1, 0.2]
gradient_boosting__max_depth	[3, 5]	[3, 5, 7, 9]
gradient_boosting__min_samples_leaf	[1, 5]	[1, 5, 10, 20]
gradient_boosting__n_estimators	[100, 200, 500]	[50, 100, 200, 500]
gradient_boosting__subsample	[0.8, 1.0]	[0.6, 0.8, 1.0]
random_forest__bootstrap	[]	[true, false]
random_forest__criterion	["gini", "entropy"]	["gini", "entropy", "log_loss"]
random_forest__max_depth	[5, 10, 20]	[null, 5, 10, 20, 30, 50]
random_forest__min_samples_leaf	[1, 4]	[1, 2, 4, 8]
random_forest__min_samples_split	[2, 10]	[2, 5, 10, 20]
random_forest__n_estimators	[100, 200, 500]	[50, 100, 200, 500]
svc__C	[0.1, 1.0, 10.0]	[0.1, 1.0, 10.0, 100.0]
svc__gamma	["scale", "auto"]	["scale", "auto"]
svc__kernel	["linear", "rbf"]	["linear", "rbf", "poly", "sigmoid"]

Halving resource: n_samples (the wizard default). This model has no top-level n_estimators resource. Use n_samples in the wizard; another estimator parameter needs an explicit compatible configuration.

Minimal training entries

Untuned:

models:
  candidate:
    model:
      type: stacking_classifier
      params: {"base_estimators": ["logistic_regression", "decision_tree"], "final_estimator": "logistic_regression", "cv": 3}

Tuned (small explicit grid):

models:
  candidate:
    model:
      type: stacking_classifier
      params: {"base_estimators": ["logistic_regression", "decision_tree"], "final_estimator": "logistic_regression", "cv": 3}
    tuning:
      strategy: grid
      metric: accuracy
      max_candidates: 4
      search_space:
        logistic_regression__C: [0.01, 0.1]

Common mistake: Confusing internal cv with the external model-selection folds, or multiplying both by a large search budget without checking the resulting fit count.

svc

Support vector classifier. Task: classification. Kernel classification for moderate-sized data where a nonlinear boundary improves on a linear baseline. Scale inputs; Core enables probability=True, which adds calibration work to training.

Sample weights: Yes; configure the workflow weight_column.

Optional package: None beyond the standard Core scikit-learn runtime.

Spark workers / serving: Not admitted by the current Spark-worker/pinned-serving model allowlist; use local inference.

Parameters and defaults

Parameter	Core catalog default	Explicit Core constructor default	Meaning
C	1.0	1.0	Regularization parameter. The strength of the regularization is inversely proportional to C.
kernel	"rbf"	'rbf'	Specifies the kernel type to be used in the algorithm.
gamma	"scale"	'scale'	Kernel coefficient for 'rbf', 'poly' and 'sigmoid'.
probability	Not a catalog control	True	Core constructor setting.

Default search spaces

Axis	Grid / halving_grid	Random / halving_random / Optuna
C	[0.1, 1.0, 10.0]	[0.1, 1.0, 10.0, 100.0]
gamma	["scale", "auto"]	["scale", "auto"]
kernel	["linear", "rbf"]	["linear", "rbf", "poly", "sigmoid"]

Halving resource: n_samples (the wizard default). This model has no top-level n_estimators resource. Use n_samples in the wizard; another estimator parameter needs an explicit compatible configuration.

Minimal training entries

Untuned:

models:
  candidate:
    model:
      type: svc
      params: {}

Tuned (small explicit grid):

models:
  candidate:
    model:
      type: svc
      params: {}
    tuning:
      strategy: grid
      metric: accuracy
      max_candidates: 4
      search_space:
        C: [0.1, 1.0]

Common mistake: Assuming the upstream SVC probability=False default is also the Core default, or underestimating fitting cost on large data.

voting_classifier

Voting ensemble classification. Task: classification. Combines independently fitted base learners. Regression averages predictions; classification can average probabilities (soft) or use a majority label vote (hard). weights follow the order of base_estimators.

Sample weights: Yes for the default configuration; every selected child and final estimator must support it.

Optional package: None beyond the standard Core scikit-learn runtime; add xgboost or lightgbm when selected as a member.

Spark workers / serving: Not admitted by the current Spark-worker/pinned-serving model allowlist; use local inference.

Parameters and defaults

Parameter	Core catalog default	Explicit Core constructor default	Meaning
base_estimators	["random_forest", "logistic_regression", "gradient_boosting"]	Library/resolver	Base learners combined by the ensemble.
voting	"soft"	'soft'	Soft averages predicted probabilities (needs predict_proba on every base model); hard takes the majority class vote.
n_jobs	1	Library/resolver	Number of base models fit in parallel. 1 = sequential, -1 = use all CPU cores. Speeds up fitting when you have several independent base learners.
calibrate_base_models	false	Library/resolver	Wrap each base classifier in CalibratedClassifierCV so its predicted probabilities are well-calibrated. Improves soft voting and stacking when base models give over/under-confident probabilities.
calibration_method	"sigmoid"	Library/resolver	Sigmoid (Platt scaling) is robust on small data; isotonic is more flexible but needs more samples to avoid overfitting.
calibration_cv	3	Library/resolver	Cross-validation folds used to fit each base model's calibrator.

Default search spaces

These axes use the catalog's default members shown above. The wizard expands your selected members instead; it keeps composition, voting and internal CV fixed. Calibrated members add estimator__ after their name.

Axis	Grid / halving_grid	Random / halving_random / Optuna
gradient_boosting__learning_rate	[0.05, 0.1, 0.2]	[0.01, 0.05, 0.1, 0.2]
gradient_boosting__max_depth	[3, 5]	[3, 5, 7, 9]
gradient_boosting__min_samples_leaf	[1, 5]	[1, 5, 10, 20]
gradient_boosting__n_estimators	[100, 200, 500]	[50, 100, 200, 500]
gradient_boosting__subsample	[0.8, 1.0]	[0.6, 0.8, 1.0]
logistic_regression__C	[0.01, 0.1, 1.0, 10.0]	[0.001, 0.01, 0.1, 1.0, 10.0, 100.0]
logistic_regression__max_iter	[500, 2000]	[100, 200, 500, 1000, 2000, 5000]
logistic_regression__penalty	["l1", "l2"]	["l1", "l2"]
logistic_regression__solver	["saga"]	["saga"]
random_forest__bootstrap	[]	[true, false]
random_forest__criterion	["gini", "entropy"]	["gini", "entropy", "log_loss"]
random_forest__max_depth	[5, 10, 20]	[null, 5, 10, 20, 30, 50]
random_forest__min_samples_leaf	[1, 4]	[1, 2, 4, 8]
random_forest__min_samples_split	[2, 10]	[2, 5, 10, 20]
random_forest__n_estimators	[100, 200, 500]	[50, 100, 200, 500]

Halving resource: n_samples (the wizard default). This model has no top-level n_estimators resource. Use n_samples in the wizard; another estimator parameter needs an explicit compatible configuration.

Minimal training entries

Untuned:

models:
  candidate:
    model:
      type: voting_classifier
      params: {"base_estimators": ["logistic_regression", "decision_tree"]}

Tuned (small explicit grid):

models:
  candidate:
    model:
      type: voting_classifier
      params: {"base_estimators": ["logistic_regression", "decision_tree"]}
    tuning:
      strategy: grid
      metric: accuracy
      max_candidates: 4
      search_space:
        logistic_regression__C: [0.01, 0.1]

Common mistake: Using full registry IDs as member names: use ridge or random_forest, for example, rather than ridge_regression or random_forest_regressor. Soft voting requires probabilities from every member.

xgboost_classifier

XGBoost classification. Task: classification. An optional gradient-boosted tree model with row/column sampling and regularization controls. Useful for tabular interactions when the matching xgboost package is available in every runtime.

Sample weights: Yes; configure the workflow weight_column.

Optional package: xgboost.

Spark workers / serving: Not admitted by the current Spark-worker/pinned-serving model allowlist; use local inference.

Parameters and defaults

Parameter	Core catalog default	Explicit Core constructor default	Meaning
n_estimators	100	100	Number of gradient boosted trees.
max_depth	6	6	Maximum tree depth for base learners.
learning_rate	0.3	0.3	Boosting learning rate (eta).
subsample	1.0	0.8	Subsample ratio of the training instances.
colsample_bytree	1.0	0.8	Subsample ratio of columns when constructing each tree.
min_child_weight	1	Library/resolver	Minimum sum of instance weights in a child. Higher values regularize against overfitting.
gamma	0.0	Library/resolver	Minimum loss reduction required to make a split. Higher = more conservative tree growth.
reg_alpha	0.0	Library/resolver	L1 regularization term on leaf weights.
reg_lambda	1.0	Library/resolver	L2 regularization term on leaf weights. Default sklearn XGBoost is 1.
random_state	42	Library/resolver	Seed for row/column subsampling when subsample or colsample_bytree < 1.0.
class_weight	null	Library/resolver	Automatically adjusts weights inversely proportional to class frequencies in the training data, useful for imbalanced targets. XGBoost has no native support for this, so it is applied as a computed sample_weight at fit time.
n_jobs	Not a catalog control	-1	Core constructor setting.

Default search spaces

Axis	Grid / halving_grid	Random / halving_random / Optuna
colsample_bytree	[]	[0.6, 0.8, 1.0]
learning_rate	[0.05, 0.1, 0.2]	[0.01, 0.05, 0.1, 0.3]
max_depth	[3, 5, 7]	[3, 5, 7, 9]
min_child_weight	[1, 3, 5]	[1, 3, 5, 7]
n_estimators	[100, 200, 500]	[100, 200, 500, 1000]
reg_alpha	[]	[0.0, 0.01, 0.1, 1.0]
reg_lambda	[0.1, 1.0, 5.0]	[0.1, 1.0, 5.0, 10.0]
subsample	[0.8, 1.0]	[0.6, 0.8, 1.0]

Halving resource: n_samples (the wizard default). This model also exposes n_estimators; selecting it requires explicit max_resources and removing that axis from the search space.

Minimal training entries

Untuned:

models:
  candidate:
    model:
      type: xgboost_classifier
      params: {}

Tuned (small explicit grid):

models:
  candidate:
    model:
      type: xgboost_classifier
      params: {}
    tuning:
      strategy: grid
      metric: accuracy
      max_candidates: 4
      search_space:
        n_estimators: [100, 200]

Common mistake: Installing xgboost only on the training driver, or assuming every classifier/booster configuration is admitted for Spark workers.

Keeping this catalog current

The explanatory text and examples are maintained here; marked parameter/search regions are regenerated from Core. From a repository checkout, run:

python skyulf-core/templates/databricks/build_model_catalog.py
python skyulf-core/templates/databricks/build_model_catalog.py --check

The generator needs only Python's standard library. Its check and the model-card coverage/example tests detect drift without needing workspace credentials.