TinyGSM: Achieving >80% on GSM8k with small language modelsarxiv.org 2 pointsconvexstrictly2 years ago1 commentSaveHideCopy link On HNComments−convexstrictlyOP2y"... we find that a duo of a 1.3B generation model and a 1.3B verifier model can achieve 81.5% accuracy, outperforming existing models that are orders of magnitude larger."
Comments
"... we find that a duo of a 1.3B generation model and a 1.3B verifier model can achieve 81.5% accuracy, outperforming existing models that are orders of magnitude larger."