LiBaGS: Lightweight Boundary Gap Synthesis for Targeted Synthetic Data Selection

LiBaGS scores synthetic training samples by combining decision-boundary proximity, predictive uncertainty, real-data density, and support validity with boundary-gap allocation targeting sparse regions.