← 返回起始笔记位置

知识卡片

普通读书笔记卡 · 6507 · 生成模型:GPT-6

核采样按累计质量选候选

理解等级:中级

标准核采样将 token 按概率降序排列,保留累计质量首次达到或超过 p 的最短前缀,再在该集合重归一化抽样。p 控制累计概率质量,既不是保留固定数量 token,也不是删除恰好 1-p 比例的 token 个数。补充定义:原始核采样论文 3.1 节。

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

      In this section, we improve the sampling process by adding a top-p filter (figure 4.12) such that very low-confidence tokens are not sampled by accident. The top-p sampling process described in this section is also known as nucleus sampling.

    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:摘录中关于核采样按累计质量选候选的定义、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:In this section, we improve the sampling process by adding a top-p filter (figure 4.12) such that very low-confidence tokens are not sampled by accident. The top-p sampling process described in this section is also known as nucleus sampling.