Synthetic Eggs in Many Baskets: The Impact of Synthetic Data Diversity on LLM Fine-Tuning

TL;DR

The study investigates how the diversity of synthetic data sources used for fine-tuning large language models affects distributional properties, robustness, and biases. By systematically varying single-source versus multi-source synthetic data across a range of Llama-3.1 target sizes and source-model scales, the authors quantify effects on perplexity, lexical/semantic diversity, vocabulary growth, and downstream safety without human data embargoes. Key findings show that multi-source synthetic data mitigates distribution collapse and improves modeling of human text, but data diversity interacts with model size to shape adversarial vulnerability and self-preference bias, with human data most effectively reducing self-bias. The work provides practical guidance for synthetic data strategies, highlighting trade-offs between diversity, output quality, and safety across model scales and languages, and identifying directions for extending analysis to multi-turn interactions and non-English data. Overall, this paper clarifies when synthetic data diversity helps or hurts LLM fine-tuning, informing safer and more effective data-generation practices.

Abstract

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.

Figures (13)

• Figure 1: Dataset augmentation process, which was repeated for three model sizes (small, medium, large). For the single-source condition, the source and target model are the same architecture. For the multi-source condition, the target architecture is one of several sources.
• Figure 2: Heaps' Law fitted curves for $V(n) = K \cdot n^{\beta}$, with $V=\text{vocabulary size}$, $n=\text{number of tokens}$, and $K$ and $\beta$ being fitted parameters.
• Figure 3: Perplexity scores of single-source, multi-source, human-source, and vanilla models on the Dolly-15k test set for Llama-small and Llama-medium.
• Figure 4: Distribution of Quality and Harmfulness ratings for Llama-8B models. Each pie chart represents the proportion of different model types (Single-Source, Multi-Source, Human-Source, and Vanilla) at each quality/harmfulness coordinate. The size of each pie chart is proportional to the total number of responses at that coordinate. The most dangerous outputs can be assumed to be located in the top-right corner (high quality + high harmfulness), labeled 'Danger Zone'.
• Figure 5: Composition of the danger zone for Llama-70B across different sizes of fine-tuning generator models.