Configuration

This page documents the configuration schema consumed by SkyulfPipeline and FeatureEngineer (the skyulf-core library).

Running the full platform? See the Backend Configuration Reference for all .env / environment variable settings (database, security, Celery, uploads, etc.).

Pipeline config

SkyulfPipeline expects:

{
  "preprocessing": [ ... ],
  "modeling": { ... }
}

Preprocessing config

The preprocessing list is executed in order.

Each step is:

{
  "name": "step_name",
  "transformer": "TransformerType",
  "params": { ... }
}

TransformerType is a string key resolved via the NodeRegistry. For the full list and per-node parameters, see:

• Reference → Preprocessing Nodes
• Reference → API → Preprocessing → pipeline

Minimal examples

# Split to avoid leakage
{"name": "split", "transformer": "TrainTestSplitter", "params": {"test_size": 0.2, "random_state": 42, "target_column": "target"}}

# Impute missing numeric values
{"name": "impute", "transformer": "SimpleImputer", "params": {"strategy": "mean", "columns": ["age"]}}

# Encode categoricals
{"name": "encode", "transformer": "OneHotEncoder", "params": {"columns": ["city"], "drop_original": True, "handle_unknown": "ignore"}}

# Scale numeric columns
{"name": "scale", "transformer": "StandardScaler", "params": {"auto_detect": True}}

Modeling config

SkyulfPipeline supports the following model types via the NodeRegistry.

Classification (9 models)

Key	Algorithm
logistic_regression	Logistic Regression
random_forest_classifier	Random Forest Classifier
svc	Support Vector Classifier
k_neighbors_classifier	K-Nearest Neighbors Classifier
decision_tree_classifier	Decision Tree Classifier
gradient_boosting_classifier	Gradient Boosting Classifier
adaboost_classifier	AdaBoost Classifier
xgboost_classifier	XGBoost Classifier (requires skyulf-core[modeling-xgboost])
gaussian_nb	Gaussian Naive Bayes

Regression (11 models)

Key	Algorithm
linear_regression	Linear Regression
ridge_regression	Ridge Regression
lasso_regression	Lasso Regression
elasticnet_regression	ElasticNet Regression
random_forest_regressor	Random Forest Regressor
svr	Support Vector Regressor
k_neighbors_regressor	K-Nearest Neighbors Regressor
decision_tree_regressor	Decision Tree Regressor
gradient_boosting_regressor	Gradient Boosting Regressor
adaboost_regressor	AdaBoost Regressor
xgboost_regressor	XGBoost Regressor (requires skyulf-core[modeling-xgboost])

Meta

Key	Purpose
hyperparameter_tuner	Wraps any model above with grid, random, Optuna, or halving search

Example:

{
  "type": "random_forest_classifier",
  "node_id": "model_node",
  "params": {
    "n_estimators": 200,
    "max_depth": 10
  }
}

Tuner example:

{
  "type": "hyperparameter_tuner",
  "base_model": {"type": "logistic_regression"},
  "strategy": "optuna",
  "search_space": {"C": [0.1, 1.0, 10.0]},
  "n_trials": 25,
  "metric": "accuracy"
}

Reproducibility and seeds

Seeding has exactly one owner: skyulf.types.DEFAULT_RANDOM_STATE (currently 42). Every stochastic component (model fitting, CV folds, tuning search, splitting, encoding cross-folds) falls back to it, so runs are reproducible out of the box. Explicit configuration always wins, in this precedence order:

1. Node params — "params": {"random_state": 7} on a model/preprocessing node controls that node only. Pass "random_state": null to opt out of seeding a specific estimator.
2. Tuner config — random_state and cv_random_state on the hyperparameter_tuner config seed the search, the CV folds, and the final refit.
3. DEFAULT_RANDOM_STATE — the shared fallback, injected at model construction.

Two exceptions, both fixed and reproducible:

• Iterative imputation defaults its seed to 0 (sklearn's own default for IterativeImputer), not 42. The canvas shows the seed in the imputation settings; set it to 42 if you want every node aligned.
• Internal helpers (feature-selection importance estimators, target/WoE encoding CV folds, profiling analyzers) carry fixed seeds that are not user-configurable — they are implementation details, not pipeline choices.

# Pin one seed for the whole tuning run:
{
  "type": "hyperparameter_tuner",
  "base_model": {"type": "random_forest_classifier"},
  "strategy": "random",
  "search_space": {"n_estimators": [100, 200]},
  "random_state": 123,
  "cv_random_state": 123
}

Where to set seeds in the canvas (per node)

Node	Control	Where in the UI	What it seeds
Train/Test Split	Random State	Split settings	Which rows go to train/test/validation
Classification, Regression, Text Classification (basic mode)	Random State	Hyperparameters → Customize	The model's own randomness (bootstrap sampling, tree splits, stochastic solvers, initialization)
Classification, Regression, Text Classification (advanced mode)	Random State	Tuning Strategy section	The tuning search and the final refit
Any training/ensemble node	Fold Split Seed	Cross Validation section	How rows are dealt to CV folds (only when shuffling is on)
Ensemble	Random State	Hyperparameter Tuning section	The ensemble tuning search and refit
Segmentation (K-Means, Mini-Batch K-Means, Gaussian Mixture)	Random State	Hyperparameters → Customize	Centroid/component initialization
Imputation (Iterative only)	Random State	Imputation settings (default 0)	The iterative imputer's estimator
Resampling (over/undersampling)	Random State	Resampling settings	Synthetic sample generation / sampling

Notes:

• Basic mode: the seed appears in the model's hyperparameter list once you switch on Customize; leaving it untouched keeps the default 42.
• Tuning never searches over seeds. The seed is a fixed control, not a search-space candidate — same seed + same data = identical tuning outcome.
• Models without a seed control are deterministic by construction: Linear Regression, K-Nearest Neighbors, Naive Bayes family, SVC/SVR, BIRCH.
• Ensemble base learners (voting/stacking) are seeded automatically at construction; per-base-learner seeds can be supplied via base_estimator_params (API) or by wiring in a model node with a customized Random State.
• Opt-out: in a hand-written config, "random_state": null leaves that estimator unseeded (run-to-run variation). The canvas always sends a number.

See "Modeling Nodes" in Reference for details.