WizardLM

WizardLM, from Microsoft Research (paper arXiv:2304.12244, April 2023), introduced Evol-Instruct: a method that uses LLMs themselves to automatically generate increasingly complex instruction data, replacing expensive human annotation. Models in 7B, 13B, 30B, and 70B sizes were released through mid-2023, fine-tuned on LLaMA bases.

At release, WizardLM-13B topped open-source leaderboard rankings on AlpacaEval and MT-Bench, and its sibling WizardCoder achieved state-of-the-art HumanEval scores among open models. Evol-Instruct became one of the most-copied training recipes in the open-model world. The catch: built on LLaMA 1, the original weights carry a non-commercial research license, so verify the specific checkpoint's terms before any commercial use.