generalfactneutralData scaling from 976 to 11,101 samples yields +20.6% improvement in tongue diagnosis model performanceComputer Vision02 Aug 2026http://arxiv.org/abs/2607.28148v1