Selecting and swapping foundation models

Select an exact model family and task through the service catalogue. A listed family can have a classifier ready while its regressor is unavailable. Check the requested task_profiles entry for status, execution_available, version, configuration and limits before fitting. Worker availability can change after inspection; typed service errors remain authoritative at submission time.

from welt import Client, Classifier

with Client() as client:  # WELT_BASE_URL and WELT_API_KEY configured locally
    entries = {entry["id"]: entry for entry in client.models()}
    for model_id in ("tabicl-v2", "kumo-medium", "tabdpt-1.3", "mitra-v2"):
        profile = next((p for p in entries[model_id]["task_profiles"]
                        if p["task"] == "classification"), None)
        if not profile or profile["status"] != "available" or not profile["execution_available"]:
            raise RuntimeError(f"Requested classifier unavailable: {model_id}")

# X_train, y_train and X_query are your consistently prepared table/split.
# Fit a new immutable predictor for each model, even on the same training table.
models = {
    model_id: Classifier(model=model_id, random_state=42).fit(X_train, y_train)
    for model_id in ("tabicl-v2", "kumo-medium", "tabdpt-1.3", "mitra-v2")
}
points = {model_id: estimator.predict(X_query) for model_id, estimator in models.items()}

Use Regressor with the same pattern and inspect regression profiles first. Named DataFrames preserve fitted columns and allow safe reordering. Existing predictors retain their exact model/configuration/seed identities; changing the model parameter requires a new fit. Reopening by predictor ID reuses that pinned identity and never switches models. Predictors retain support/preprocessing state when idle native caches are evicted.

Comparing these predictions is not a benchmark. Evaluate on the same appropriate held-out rows and metric, keep preprocessing inside training folds, and separate model selection from unbiased assessment. Automatic search, enrichment, conformal coverage and larger workloads remain separate capabilities. Native classifier probabilities and regression mean points do not promise calibrated confidence.

Mitra-v2 uses an explicit zero-shot recipe without prediction-time fine-tuning. Its native regression histogram mean is a point in the original target units; it does not provide a calibrated interval or publisher benchmark guarantee.