Teaching Large Language Models to Self-Debug

Benchmark Model Rank Results
code-generation-on-mbppGPT-4 (Self-Debugging with unit tests + trace)#32Accuracy: 80.2
code-generation-on-mbppGPT-3.5 Turbo (Self-Debugging with unit tests + trace)#49Accuracy: 72.8
code-generation-on-mbppcode-davinci-002 175B (Self-Debugging with unit tests + trace)#54Accuracy: 70.8
code-generation-on-mbppGPT-3.5 Turbo (3-shot)#65Accuracy: 67.6
code-generation-on-mbppcode-davinci-002 175B (3-shot)#83Accuracy: 61.4
code-generation-on-mbppStarCoder 15.5B (Self-Debugging with unit tests + trace)#100Accuracy: 53.2
code-generation-on-mbppStarCoder 15.5B (3-shot)#117Accuracy: 47.2