Read model outputs

Main artifacts

• Training_Set_Detailed_*.csv: modeling-pool features and SHAP.

• Feature_<name>: descriptor value.

• SHAP_<name> / SHAP_Std_<name>: attribution mean/dispersion across available validation appearances.

• Test_Set_Detailed_*.csv: explicit held-out samples, if provided.

• Realtest_Pred / Realtest_Pred_Std: original ensemble prediction and between-model dispersion.

• Target: supplied target, not a new measurement.

The engine uses repeated random splits. With too few repeats some samples may never appear in validation; legacy aggregation can then contain zero-filled values. Check validation coverage before reporting metrics. Ensemble spread is not a calibrated confidence interval.

Automatic feature selection is exploratory, not a nested-CV implementation of the separately frozen manuscript benchmark. Predefine features, or evaluate the entire selection procedure inside training folds for scientific comparisons. Keep permanent tests out of selection, explanation ranking and parameter choice.

SHAP values are model attributions, not measured effects or candidate scores. Open actual files with From model attribution to chemical meaning.