Synthetic Eggs in Many Baskets: The Impact of Synthetic Data Diversity on LLM Fine-Tuning
Max Schaffelder, Albert Gatt · 2025
Read Synthetic Eggs in Many Baskets: The Impact of Synthetic Data Diversity on LLM Fine-Tuning by Max Schaffelder, Albert Gatt (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-01490.txt.
Description
Max Schaffelder and Albert Gatt ask what happens when a large language model is fine-tuned on text that other models generated. Their study tests whether drawing that synthetic training data from many different sources, rather than one, changes how the fine-tuned model behaves.
Three effects are measured: distribution collapse, robustness against attacks, and self-preference bias, meaning a model favoring its own outputs. Diverse sources eased collapse and kept the model's outputs broad. Both human and synthetic data could strip away safeguards, though synthetic data gave higher-quality outputs, which the authors note makes them more usable and more dangerous. Fine-tuning also cut self-preference bias, with human data working best.
Catalog record
| CREATOR | Max Schaffelder, Albert Gatt |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 76 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2511-01490