notesum.ai

Published at December 6

Neuro-Symbolic Data Generation for Math Reasoning

cs.AI

Released Date: December 6, 2024

Authors: Zenan Li¹, Zhi Zhou¹, Yuan Yao¹, Yu-Feng Li¹, Chun Cao¹, Fan Yang², Xian Zhang², Xiaoxing Ma¹

Aff.: ¹State Key Lab of Novel Software Technology, Nanjing University, China; ²Microsoft Research Asia

Arxiv: http://arxiv.org/pdf/2412.04857v1

Refer to caption

Model	#Dataset	LLaMA-2 7B Base		LLaMA-2 13B Base		Mistral 7B Base
Model	#Dataset	GSM8K	MATH	GSM8K	MATH	GSM8K	MATH
WizardMath	$>$ 240K	54.9	10.7	63.9	14.0	83.2	33.0
MuggleMATH	157K	68.4	8.4	74.0	9.4	-	-
MAmmoTH^†	260K	50.5	10.4	56.3	12.9	61.9	17.5
MetaMath	395K	66.5	19.8	72.3	22.4	77.7	28.2
Ours	860K	79.0	30.4	84.1	33.7	86.8	37.3
$\Delta$		$\uparrow$ 10.6	$\uparrow$ 10.6	$\uparrow$ 10.1	$\uparrow$ 11.3	$\uparrow$ 3.6	$\uparrow$ 4.3