Be like a Goldfish, Don't Memorize! Mitigating Memorization in Generative LLMs
Abstract
Large language models can memorize and repeat their training data, causing privacy and copyright risks. To mitigate memorization, we introduce a subtle modification to the next-token training objective that we call the goldfish loss. During training, a randomly sampled subsets of tokens are excluded from the loss computation. These dropped tokens are not memorized by the model, which prevents verbatim reproduction of a complete chain of tokens from the training set. We run extensive experiments training billion-scale LLaMA-2 models, both pre-trained and trained from scratch, and demonstrate significant reductions in extractable memorization with little to no impact on downstream benchmarks.
Cite
Text
Hans et al. "Be like a Goldfish, Don't Memorize! Mitigating Memorization in Generative LLMs." ICLR 2025 Workshops: Data_Problems, 2025.Markdown
[Hans et al. "Be like a Goldfish, Don't Memorize! Mitigating Memorization in Generative LLMs." ICLR 2025 Workshops: Data_Problems, 2025.](https://mlanthology.org/iclrw/2025/hans2025iclrw-like/)BibTeX
@inproceedings{hans2025iclrw-like,
title = {{Be like a Goldfish, Don't Memorize! Mitigating Memorization in Generative LLMs}},
author = {Hans, Abhimanyu and Wen, Yuxin and Jain, Neel and Kirchenbauer, John and Kazemi, Hamid and Singhania, Prajwal and Singh, Siddharth and Somepalli, Gowthami and Geiping, Jonas and Bhatele, Abhinav and Goldstein, Tom},
booktitle = {ICLR 2025 Workshops: Data_Problems},
year = {2025},
url = {https://mlanthology.org/iclrw/2025/hans2025iclrw-like/}
}