Abstract

Code Large Language Models (Code LLMs), such as StarCoder, have demonstratedexceptional performance in code-related tasks. However, most existing modelsare solely pre-trained on extensive raw code data without instructionfine-tuning. In this paper, we introduce WizardCoder, which empowers Code LLMswith complex instruction fine-tuning, by adapting the Evol-Instruct method tothe domain of code. Through comprehensive experiments on four prominent codegeneration benchmarks, namely HumanEval, HumanEval+, MBPP, and DS-1000, weunveil the exceptional capabilities of our model. It surpasses all otheropen-source Code LLMs by a substantial margin. Moreover, our model evenoutperforms the largest closed LLMs, Anthropic's Claude and Google's Bard, onHumanEval and HumanEval+. Our code, model weights, and data are public athttps://github.com/nlpxucan/WizardLM