List your company

Mathematical Reasoning on GSM8K (Custom Sampling Metrics)

66.4m1@t1Self-Refine
8.26423.35738.4553.543Jul 25, 2024

Evaluation results

MethodLinks
Self-RefineBackbone=GPT-3.52024.0766.480.2——6171.6
GLoREComponent=ORM2024.0748.2———49.557.1
GLoREComponent=SORM2024.0748.2———51.659.7
GLoREComponent=Direct2024.0748.2———47.459.2
RISEBackbone=Llama2 Base, Iteration=22024.0737.3515568.4——
Self-RefineBackbone=Llama2 Base, Iteration=22024.0737.350.5——33.344.5
Eurus-7B-SFTBackbone=Eurus-7B2024.0736.366.347.953.1——
Self-RefineBackbone=Eurus-7B-SFT2024.0736.365.9——26.242.8
RISEBackbone=Llama2 Base, Iteration=12024.0735.649.750.763.9——
RISEBackbone=Mistral-7B, Iteration=12024.0735.350.659.268.6——
RISEBackbone=Mistral-7B2024.0733.749.43946.9——
Self-RefineBackbone=Mistral-7B2024.0733.748.5——21.237.9
RISEBackbone=Llama2 Base, Stage=Boost2024.0732.945.439.255.5——
SFT on oracle dataBackbone=Llama2 Base, Data=Only correct data2024.0727.442.23443.6——
SFT on oracle dataBackbone=Llama2 Base, Data=Correct and incorrect2024.0725.741.831.241.5——
RISEBackbone=Llama2 Base2024.0710.522.811.113.9——
Self-RefineBackbone=Llama2 Base2024.0710.522.4——7.113