Predict a number

Classification predicts a label; regression predicts a numeric target. Begin with the complete Yacht walkthrough: download the real research table, print its rows/schema, split train/test, fit a Regressor, predict the held-out rows, and calculate MAE, RMSE and R².

# Continue after preparing the Yacht train/test tables in Start.
from welt import Regressor
from sklearn.metrics import mean_absolute_error, root_mean_squared_error, r2_score

model = Regressor(model="tabicl-v2", random_state=9)
model.fit(train, target=target)
predictions = model.predict(query)
print("Test MAE:", mean_absolute_error(y_test, predictions))
print("Test RMSE:", root_mean_squared_error(y_test, predictions))
print("Test R²:", r2_score(y_test, predictions))

MAE/RMSE use recorded target units; smaller is better. R² can be negative. Compare actual and predicted rows to understand misses. These are held-out measurements, not a publisher benchmark or guarantee. Use a validation split if tuning.

A task-specific available regression profile is required; no task/model fallback occurs. predict returns finite target-unit mean points, not class probabilities, calibrated intervals or conformal coverage. Predictors keep their original task, version and configuration. The notebook teaches the full recipe.