Code Llama: Open Foundation Models for Code

Benchmark Model Rank Results
code-generation-on-mbppCode Llama - Python 70B (3-shot)#72Accuracy: 65.5
code-generation-on-mbppCode Llama 70B (3-shot)#79Accuracy: 62.4
code-generation-on-mbppCode Llama - Instruct 70B (3-shot)#81Accuracy: 62.2
code-generation-on-mbppUnnatural Code Llama 34B (3-shot)#85Accuracy: 61.2
code-generation-on-mbppCode Llama - Instruct 34B (3-shot)#93Accuracy: 57
code-generation-on-mbppCode Llama - Python 34B (3-shot)#95Accuracy: 56.2
code-generation-on-mbppCode Llama 34B (3-shot)#99Accuracy: 55
code-generation-on-mbppGPT-3.5 Turbo#103Accuracy: 52.2
code-generation-on-mbppCode Llama - Instruct 13B (3-shot)#107Accuracy: 49.4
code-generation-on-mbppCode Llama - Python 13B (3-shot)#110Accuracy: 49
code-generation-on-mbppCode Llama - Python 7B (3-shot)#114Accuracy: 47.6
code-generation-on-mbppCode Llama 13B (3-shot)#119Accuracy: 47
code-generation-on-mbppCode Llama - Instruct 7B (3-shot)#126Accuracy: 44.4
code-generation-on-mbppCode Llama 7B (3-shot)#133Accuracy: 41.4
code-generation-on-humanevalCode Llama - Instruct 70B#89Pass@1: 67.8
arithmetic-reasoning-on-gsm8kCode Llama - Python 34B#254Accuracy: 34.420