LWCal Calibrates Tabular Classifiers Using Noisy Labels

LWCal provides a CPU-only post-hoc calibration method for tabular models that operates without clean validation labels or retraining. It down-weights calibration examples where the base model's held-out probability contradicts the label, including a Gated-LWCal variant that reverts to raw scores during extreme inconsistency.