← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 6499.h · 生成模型:GPT-6

采样多样性与扩展预算分开

理解等级:中级

温度和 top-p 本身修改单步选择分布;增加候选数量或步骤才显式扩展推理预算。采样也可能间接改变回答长度,所以实验需同时记录实际 token 与时间,避免把所有准确率变化都归因于“多算了”。

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

      not inference-time scaling techniques themselves. We can also see that the runtime increased from 10.1 min to 30.7 min. This is not due to the sampling code overhead, but rather because the model now generates longer responses in some cases.

    参考来源

    - 位置:《Build a Reasoning Model (From Scratch)》第04章《推理时计算扩展》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Build a Reasoning Model (From Scratch)/_epub-src/OEBPS/Text/chapter-4.html) - 结论依据:摘录中关于采样多样性与扩展预算分开的定义、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:not inference-time scaling techniques themselves. We can also see that the runtime increased from 10.1 min to 30.7 min. This is not due to the sampling code overhead, but rather because the model now generates longer responses in some cases.