WizardLM-70B is a large language model developed by the WizardLM team, utilizing the Evol-Instruct method to enhance instruction-following capabilities. Based on the Llama 2 architecture, it was trained on a sophisticated dataset generated by an evolutionary algorithm that iteratively increases the complexity of instructions. This process enables the model to handle intricate, multi-step tasks and logical reasoning more effectively than models trained on simpler instruction sets. The 70-billion parameter model represents a significant scale in its release series, offering improved performance on benchmarks like MT-Bench and AlpacaEval. By leveraging the base Llama 2 weights, it maintains the fundamental linguistic capabilities of the foundation model while specializing in complex prompt adherence and open-domain conversation.