Train a model

Fit candidate models on top of your cleaning pipeline, ranked by the metric your goal calls for — with no train/test leakage.

Data source

Uses the raw data and the approved recipe from steps 1–6. The recipe is replayed per fold, so this path is leakage-safe by construction.

Trusted-input mode: PrePro Auto did not see the raw data or recipe, so leakage from prior all-rows fitting can’t be verified. We add no target encoding of our own and scan for obvious leaks.

Setup

Choose from your current/updated columns.

Plain English. “Catch fraud” → PR-AUC; “estimate price” → RMSE; “score/rank” → ROC-AUC.

Auto-detected on “Suggest”. Switch to filter the model list.

Recommendation

Problem—
Ranking metric—
Class balance—

The recommended models are pre-checked below. Add or remove any model you like, then train.

Choose models · search and pick any — or leave all unchecked to train the recommended defaults

No models selected — training will use the recommended defaults.

Tuning & options

Search strategies use scikit-learn, which ships with pip install prepro-auto — nothing extra to install. Optuna (Bayesian) needs pip install optuna. The boosted-tree models (XGBoost / LightGBM / CatBoost) and ONNX export need pip install "prepro-auto[ml]".

Hold out the latest rows — no future leaks into training.

Keep each entity (customer/store) entirely in train or test.

Leaderboard · cross-validated on the training split, then scored once on held-out test

Test it

Enter values for your dataset's feature columns (these are the columns of your current cleaned data, minus the target). They run through the full pipeline — exactly what production would do.

Export

Every artifact bundles preprocessing and the model. The model is never shipped on its own.