🚀Major Update: Introducing WizardLM 13B Version.

• On difficulty-balanced Evol-Instruct testset, evaluated by GPT-4: WizardLM-13B achieves 89.1% of ChatGPT, Vicuna-13B achieves 86.9%, and WizardLM-7B achieves 78%.
• The 13B version is trained on instruction data evolved from real-world human conversations (ShareGPT), while the 7B version is trained on instruction data evolved from machine-generated data (Alpaca).
• We provide a comparison between the performance of the WizardLM-13B and ChatGPT on different skills to establish a reasonable expectation of WizardLM's capabilities.