How to Synthesize Text Data without Model Collapse?

Model collapse in synthetic data indicates that iterative training on self-generated data leads to a gradual decline in performance. With the proliferation of AI models, synthetic data will fundamentally reshape the web data ecosystem. Future GPT-{n} models will inevitably be trained on a blend of synthetic and humanproduced data. In this paper, we focus on two questions: what is the impact of synthetic data on language model training, and how to synthesize data without model collapse? We further conduct statistical analysis on synthetic data to uncover distributional shift phenomenon and over-concentration of n-gram features. Inspired by the above findings, we propose token editing on human-produced data to obtain semi-synthetic data. 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 conduct extensive experiments on pre-training from scratch, continual pre-training, and supervised finetuning. The results validate our theoretical proof that token-level editing improves data quality and enhances model performance. As generative artificial intelligence (AI) (Rombach et al., 2021; Achiam et al., 2023) becomes increasingly prevalent in research and industry, synthetic data will proliferate throughout the web data ecosystem. Consequently, future training of GPT-{n} on a mixture of synthetic and humanproduced data will be inevitable. Thus, model collapse is a critical concern that must be considered when training models on synthetic data. Model collapse refers to a degenerative process in which the output data of learned generative models contaminates the training sets of subsequent generations. As shown in Figure 1, iterative training coupled with data synthesis induces a progressive accumulation of test errors (Shumailov et al., 2024; Dohmatob et al., 2024a). Consequently, generative models increasingly overfit to synthetic data distributions, failing to capture the complexity in human-produced data.