WizardCoder

WizardCoder, from Microsoft and Hong Kong Baptist University researchers in June 2023, adapted the Evol-Instruct method to code. Starting from the open StarCoder 15B base, the team used code-specific evolutionary prompts, adding debugging and time-space complexity constraints, to generate a harder and more diverse instruction dataset, then fine-tuned StarCoder on it.

The 15B model surpassed every open-source code LLM by a wide margin and even beat Anthropic's Claude and Google's Bard on HumanEval and HumanEval+, remarkable for its size; a 34B variant matched GPT-3.5 on HumanEval. Code, weights, and data were released publicly. License caveat: as a StarCoder derivative it inherits the licensing terms of its base, so verify before commercial use. A landmark in instruction-tuning for code that influenced the whole code-LLM wave of 2023.