Automatic Instruction Evolving for Large Language Models

Fine-tuning large pre-trained language models with Evol-Instruct has achievedencouraging results across a wide range of tasks. However, designing effectiveevolving methods for instruction evolution requires substantial humanexpertise. This paper proposes Auto Evol-Instruct, an end-to-end framework thatevolves instruction datasets using large language models without any humaneffort. The framework automatically analyzes and summarizes suitableevolutionary strategies for the given instruction data and iteratively improvesthe evolving method based on issues exposed during the instruction evolutionprocess. Our extensive experiments demonstrate that the best method optimizedby Auto Evol-Instruct outperforms human-designed methods on various benchmarks,including MT-Bench, AlpacaEval, GSM8K, and HumanEval.