List your company

Mathematical Reasoning on MATH 500, GSM8K, AIME 24, AMC 23, OlympiadBench, and Minerva Math (test)

46.92Final AccuracyAdaRFT
36.811239.435642.0644.6844Apr 7, 2025

Evaluation results

MethodLinks
AdaRFTModel=Qwen2.5 7B, Setup=uniform2025.0446.92234.16000
AdaRFTModel=Qwen2.5 7B, Setup=skew-difficult2025.0446.83239.92000
AdaRFTModel=Qwen2.5 7B, Setup=skew-easy2025.0445.94247.44000
PPO (w/ FC)Model=Qwen2.5 7B, Setup=skew-difficult2025.0445.54243.1222.592,188.08
PPO (w/ Filter)Model=Qwen2.5 7B, Setup=skew-difficult2025.0445.35254.2262.5256,355.5
PPOModel=Qwen2.5 7B, Setup=skew-easy2025.0445.07235.2750204,705.4
PPO (w/ Filter)Model=Qwen2.5 7B, Setup=skew-easy2025.0444.98233.1342.5173,963.21
PPOModel=Qwen2.5 7B, Setup=uniform2025.0444.7243.8232.5133,169.66
PPO (w/ Filter)Model=Qwen2.5 7B, Setup=uniform2025.0444.57263.1157.5236,051.53
PPO (w/ FC)Model=Qwen2.5 7B, Setup=uniform2025.0444.32240.6217.571,684.34
PPOModel=Qwen2.5 7B, Setup=skew-difficult2025.0444.17246.2160245,909.04
PPO (w/ FC)Model=Qwen2.5 7B, Setup=skew-easy2025.0443.46240.6657.5235,535.18
AdaRFTModel=Qwen2.5 Math 1.5B, Setup=uniform2025.0441.11121.31000
AdaRFTModel=Qwen2.5 Math 1.5B, Setup=skew-difficult2025.0440.48122.24000
AdaRFTModel=Qwen2.5 Math 1.5B, Setup=skew-easy2025.0439.18120.52000
PPOModel=Qwen2.5 Math 1.5B, Setup=skew-easy2025.0438.46121.1526.7161,938.4
PPO (w/ Filter)Model=Qwen2.5 Math 1.5B, Setup=skew-easy2025.0438.15115.1235212,417.52
PPO (w/ FC)Model=Qwen2.5 Math 1.5B, Setup=skew-easy2025.0438.09121.9958.3354,269.65
PPO (w/ Filter)Model=Qwen2.5 Math 1.5B, Setup=uniform2025.0437.87126.3586.7526,570.2
PPOModel=Qwen2.5 Math 1.5B, Setup=skew-difficult2025.0437.41132.9571.7435,716.85
PPO (w/ FC)Model=Qwen2.5 Math 1.5B, Setup=uniform2025.0437.27126.480486,067.2
PPO (w/ Filter)Model=Qwen2.5 Math 1.5B, Setup=skew-difficult2025.0437.22128.281.7496,281.8
PPO (w/ FC)Model=Qwen2.5 Math 1.5B, Setup=skew-difficult2025.0437.22130.9126.7162,094.56
PPOModel=Qwen2.5 Math 1.5B, Setup=uniform2025.0437.2126.8256.7344,311.88