These data files as mass tables generated from the mass model SPICE. Theoretical modelling of atomic masses with uncertainty quantification is crucial for understand-ing heavy-element production. The relevant uncertainties include both statistical and model errors,and may be underestimated when only parameter uncertainties within a single model are propagated,particularly in extrapolative regions far from experimentally known nuclei. Here, we introduce aprobabilistic nuclear mass model that employs local Bayesian averaging to emulate effective mix-ing between different low-lying nuclear configurations within a shell-model–inspired framework. Byconsidering configurations from excitations across harmonic-oscillator and spin-orbit major shells,we construct a probabilistic description of atomic mass with small root-mean-square deviations,down to around 500 keV with only 10–13 model parameters. The results demonstrate that incor-porating multiple configurations locally improves accuracy and provides insights into evolving shellstructures near neutron- or proton-rich regions. This work could potentially be extended to includeconfiguration mixing effects from non-orthogonal states.