← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8673.a · 生成模型:GPT-6

融合不能保证遗忘敏感信息

理解等级:高级

权重融合可能在某些实验中削弱独特行为或偏见,但也可能保留、恢复或引入其他行为。不能把“未共享信息通常弱化”升级成隐私删除保证;需单独审计目标信息与保留能力。

这张卡片看不懂?

已有反馈 · 0 条

    还没有人提交反馈。

    在整书笔记中的位置

    原书出处 · 1 处

    • · 查看原书依据

      Zaman et al. have made a very interesting observation: when you fuse models, the shared capabilities of the models are preserved, while the unshared capabilities are usually lost. This principle can be leveraged to use model fusion as a means to remove undesirable properties from LLMs.

    参考来源

    - 位置:《Designing Large Language Model Applications》第07章《进阶微调技术》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch07.html) - 结论依据:摘录中关于融合不能保证遗忘敏感信息的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。 - 原始内容:Zaman et al. have made a very interesting observation: when you fuse models, the shared capabilities of the models are preserved, while the unshared capabilities are usually lost. This principle can be leveraged to use model fusion as a means to remove undesirable properties from LLMs.