← 返回起始笔记位置

知识卡片

普通读书笔记卡 · 6620 · 生成模型:GPT-6

混合权重决定有效训练暴露

理解等级:中级

多来源数据的实际贡献取决于采样权重与训练步数,不只是各子集文件大小。上采样小而重要的来源可增加暴露,也会增加重复与过拟合风险;不存在脱离模型、预算和目标任务的唯一最佳混合比例。

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

      optimal proportions. For example, Wikipedia, academic texts, and smaller subsets were upsampled by up to three times in The Pile dataset. More involved techniques like DoReMi and RegMix are also used to calculate the right data mixture. Meta noted that for Llama 3, it empirically arrived at a data mixture where 50% of the tokens are about general knowledge, 25% are about math and reasoning,

    参考来源

    - 位置:《Designing Large Language Model Applications》第02章《预训练数据》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch02.html) - 结论依据:摘录中关于混合权重决定有效训练暴露的定义、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:optimal proportions. For example, Wikipedia, academic texts, and smaller subsets were upsampled by up to three times in The Pile dataset. More involved techniques like DoReMi and RegMix are also used to calculate the right data mixture. Meta noted that for Llama 3, it empirically arrived at a data mixture where 50% of the tokens are about general knowledge, 25% are about math and reasoning,