Unveiling the Flaws: Exploring Imperfections in Synthetic Data and Mitigation Strategies for Large Language Models

Synthetic data has been proposed to alleviate high-quality data scarcity for training large language models (LLMs). While prior work shows synthetic data can improve downstream benchmark performance, the authors identify inherent flaws: uniform formats in synthetic data induce pattern overfitting and significant output distribution shifts that degrade instruction-following ability. Focusing on synthetic question–answer (Q-A) pairs, the paper analyzes these flaws and introduces an unlearning-based mitigation strategy using a controllable lower-bounded forgetting loss. Experiments show the method reverses instruction-following degradation caused by pattern overfitting without sacrificing benchmark performance and at relatively low cost, offering guidance for robust and efficient LLM training with synthetic data.

Jie Chen, Yupeng Zhang, Bingning Wang, Wayne Xin Zhao, Ji-Rong Wen, Weipeng Chen