Synthetic Eggs in Many Baskets: The Impact of Synthetic Data Diversity on LLM Fine-Tuning
This addresses the problem of model behavior degradation for AI developers using synthetic data, though it is incremental in exploring specific dimensions like diversity.
The paper tackled the impact of synthetic data diversity on fine-tuned large language models, finding that diverse sources mitigate distribution collapse and preserve output quality, while human data is most effective at reducing self-preference bias.
As synthetic data becomes widely used in language model development, understanding its impact on model behavior is crucial. This paper investigates the impact of the diversity of sources of synthetic data on fine-tuned large language models. We focus on three key dimensions: distribution collapse, adversarial robustness, and self-preference bias. Our findings reveal that fine-tuning models on synthetic data from diverse sources can mitigate distribution collapse, preserving the breadth of the output distribution and the diversity of the output text. Furthermore, while both human and synthetic fine-tuning data can remove safeguards, the latter preserves higher output quality, thus making outputs potentially more usable and dangerous. Finally, fine-tuning reduces self-preference bias, with human data being the most effective, followed by multi-source synthetic data.