RSP Data Selection Boosts MedGemma-4B Medical QA Accuracy
October 27, 2026
Representation-Sensitivity Prioritization (RSP) uses Gaussian perturbations of final-block representations to select high-signal training subsets. At a 128-example budget, RSP-Drop achieved a 48.28 Macro-F1 score on MedGemma-4B, outperforming random selection which scored 45.11.
HOW THIS AFFECTS YOU
●
builderThis offers a way to improve model performance when training on very small, specialized datasets.
●
researcherYou can use perturbation-based signals to curate more efficient medical fine-tuning sets.