知识卡片

普通读书笔记卡 · 8673 · 生成模型:GPT-6

权重平均需要参数坐标兼容

理解等级:高级

同名同形参数不一定在不同训练模型中承担相同语义,尤其存在神经元置换等对称性。共同基础模型上的相近微调常更易比较;平均或加权前应核对架构、tokenizer与坐标对应,再评估合成结果。

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

      Averaging The simplest way to combine multiple models is to average their parameters. Simple averaging has been shown to be quite effective. Weighted averaging During averaging, certain models or even certain layers in models can be weighted more. Interpolation Each model can be weighted by a factor w1, w2,…wn, with: w1 + w2 + w3 +...wn = 1 w1p1 + w2p2 + w3p3 +...wnpn where

    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:摘录中关于权重平均需要参数坐标兼容的定义、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:Averaging The simplest way to combine multiple models is to average their parameters. Simple averaging has been shown to be quite effective. Weighted averaging During averaging, certain models or even certain layers in models can be weighted more. Interpolation Each model can be weighted by a factor w1, w2,…wn, with: w1 + w2 + w3 +...wn = 1 w1p1 + w2p2 + w3p3 +...wnpn where