Liang Yue, Yihong Tang, Kehai Chen, Jie Liu, Min Zhang
MASTER is a novel data augmentation method that uses multi-agent simulated teaching to enhance large language models' instruction-following capabilities and reasoning abilities.
The process of instruction fine-tuning is vital for improving the abilities of language models to follow instructions and perform specific tasks. However, creating high-quality data for fine-tuning large models is difficult and expensive. To tackle this, researchers developed MASTER, a new method that uses simulated conversations between multiple agents to generate rich instructional data. This approach has led to the creation of BOOST-QA, a dataset that enhances existing data, resulting in models that perform better on various tasks, especially those requiring complex reasoning.
