WizardLM is an AI-generated instruction data method that improves large language models (LLMs) by automatically creating complex instructions. The method, called Evol-Instruct, uses LLMs to evolve instructions from simple ones into more complex ones, enabling the creation of diverse and challenging instruction data. This data is then used to fine-tune LLMs, such as LLaMA, to enhance their performance on complex tasks. The Evol-Instruct process involves two main components: Instruction Evolver and Instruction Eliminator. The Instruction Evolver generates more complex instructions through in-depth evolving (adding constraints, deepening, concretizing, increasing reasoning steps, and complicating input) and in-breadth evolving (generating new instructions). The Instruction Eliminator filters out failed instructions that do not meet the required complexity or quality standards. WizardLM, a model fine-tuned using Evol-Instruct data, outperforms existing models like Alpaca, Vicuna, and ChatGPT in human evaluations on complex tasks. It achieves higher performance on high-difficulty instructions and is preferred by human annotators over ChatGPT. On GPT-4 automatic evaluations, WizardLM achieves over 90% of ChatGPT's performance on 17 out of 29 skills. The Evol-Instruct testset, which includes real-world human instructions from diverse sources, demonstrates that the generated instructions are more complex and diverse than those created by humans. WizardLM's performance on this testset shows that it can handle complex instructions more effectively than existing models. The study highlights the potential of using AI-evolved instructions to enhance LLMs, as it allows for the mass production of complex instruction data at a lower cost. However, the study also acknowledges the limitations of the current methods, including the need for further research to address ethical and societal implications of AI-evolved instructions.WizardLM is an AI-generated instruction data method that improves large language models (LLMs) by automatically creating complex instructions. The method, called Evol-Instruct, uses LLMs to evolve instructions from simple ones into more complex ones, enabling the creation of diverse and challenging instruction data. This data is then used to fine-tune LLMs, such as LLaMA, to enhance their performance on complex tasks. The Evol-Instruct process involves two main components: Instruction Evolver and Instruction Eliminator. The Instruction Evolver generates more complex instructions through in-depth evolving (adding constraints, deepening, concretizing, increasing reasoning steps, and complicating input) and in-breadth evolving (generating new instructions). The Instruction Eliminator filters out failed instructions that do not meet the required complexity or quality standards. WizardLM, a model fine-tuned using Evol-Instruct data, outperforms existing models like Alpaca, Vicuna, and ChatGPT in human evaluations on complex tasks. It achieves higher performance on high-difficulty instructions and is preferred by human annotators over ChatGPT. On GPT-4 automatic evaluations, WizardLM achieves over 90% of ChatGPT's performance on 17 out of 29 skills. The Evol-Instruct testset, which includes real-world human instructions from diverse sources, demonstrates that the generated instructions are more complex and diverse than those created by humans. WizardLM's performance on this testset shows that it can handle complex instructions more effectively than existing models. The study highlights the potential of using AI-evolved instructions to enhance LLMs, as it allows for the mass production of complex instruction data at a lower cost. However, the study also acknowledges the limitations of the current methods, including the need for further research to address ethical and societal implications of AI-evolved instructions.