ToEdit: How to Synthesize Text Data to Avoid Model Collapse?

Xuekai Zhu, Daixuan Cheng, Hengli Li, Kaiyan Zhang, Ermo Hua, Xingtai Lv, Ning Ding, Zhouhan Lin, Zilong Zheng, Bowen Zhou

Keywords: synthetic data, model collapse

Abstract:

We explore model collapse caused by synthetic data, where AI models trained on such data experience a gradual decline in performance.
Our initial analysis examines language model pretraining on mixed human and synthetic data, highlighting performance degradation. Further statistical analysis reveals distributional shifts and an over-concentration of n-gram features caused by synthetic data. Inspired by these insights, we propose token-level editing on human data, to obtain semi-synthetic data instead of fully using model outputs. As a proof of concept, we theoretically demonstrate that token-level editing can prevent model collapse, as the test error is constrained by a finite upper bound. We conducted extensive experiments on pretraining, continual pretraining, and supervised fine-tuning of language models. The results validate our theoretical proof that token-level editing improves data quality and enhances model performance.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 11413