Be like a Goldfish, Don't Memorize! Mitigating Memorization in Generative LLMs
This addresses privacy and copyright issues for users and developers of generative LLMs, representing an incremental improvement over existing methods.
The paper tackles the problem of memorization in large language models, which poses privacy and copyright risks, by introducing a goldfish loss that excludes random token subsets during training, resulting in significant reductions in extractable memorization with minimal impact on downstream benchmarks.
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, 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.