arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2406.10209cs.CL
收藏

像金鱼一样,不要记忆!缓解生成式大语言模型中的记忆问题

Be like a Goldfish, Don't Memorize! Mitigating Memorization in Generative LLMs

University of Maryland(马里兰大学);ELLIS Institute Tübingen(埃利斯研究所蒂宾根分支);Max Planck Institute for Intelligent Systems(马克斯·普朗克智能系统研究所);Tübingen AI Center(蒂宾根人工智能中心)

Abhimanyu Hans, Yuxin Wen, Neel Jain, John Kirchenbauer, Hamid Kazemi, Prajwal Singhania, Siddharth Singh, Gowthami Somepalli, Jonas Geiping, Abhinav Bhatele, Tom Goldstein

AI总结:

针对大语言模型记忆训练数据引发的隐私版权风险,提出名为goldfish loss的微调训练目标,通过随机排除部分词元的损失计算缓解记忆,在十亿级Llama-2模型上验证可显著降低可提取记忆且几乎不影响下游性能。

摘要:

大语言模型会记忆并复现其训练数据,引发隐私与版权风险。为缓解记忆问题,我们对下一词元训练目标进行了微调,将其命名为goldfish loss(金鱼损失)。训练期间,随机采样的词元子集被排除在损失计算之外。这些被丢弃的词元不会被模型记忆,从而阻止模型逐字复现训练集中的完整词元链。我们对十亿参数规模的Llama-2模型开展了大量实验,涵盖预训练模型与从头训练模型,结果表明可提取记忆量显著降低,且对下游基准测试几乎没有影响。

英文摘要:

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.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑