Refining Large Language Models with Self-Generated Data Through Iterative Training

Xiao Hu, Muxi Diao, Jizhi Zhang, Xin Chen, Xianyuan Zhan

Keywords: self-critique training, self-generated data curation, LLM alignment

Abstract: Large language models (LLMs) have seen staggering progress in recent years. Contemporary LLMs rely on an immense amount of data for training, however, as LLMs continue to advance, the availability of high-quality external data is reaching a bottleneck, highlighting the need for model-generated data for further improvement. Although promising, directly utilizing the self-generated data for model training without scrutinized assessment or filtering can easily lead to deteriorated performance, or in other words, ``garbage in, garbage out". In this study, our insight is to carefully craft a \textit{self-critique} process, by equipping the LLMs with the ability to be self-aware and discriminative to the quality of its generated data. We introduce a co-evolved self-critique framework that enables an LLM to simultaneously enhance both its generative and evaluative capabilities through an iterative training process. This provides a scalable solution to ensure high-quality self-generated data and facilitate sustained model improvement. Fine-tuning Llama-3 and Qwen2.5 models using this framework results in encouraging improvements in both instruction-following and discriminative abilities, demonstrating the effectiveness of our method.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 115