← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 6762.e · 生成模型:GPT-6

核采样按累计概率达到阈值截断

理解等级:中级

将候选按概率降序排列,保留累计概率至少达到p的最小前缀,再归一采样。不是保留累计小于p的最小集合;候选数可随分布变,随机种子有助同环境复现,但跨实现和硬件仍可能不同。

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

      Let’s try to generate text using the nucleus sampling method (listing 10.7). In this method, instead of choosing among the k most likely words, the model looks at the smallest set of words whose cumulative probability is smaller than p with the top_p parameter. So if there are only a few candidates with large probabilities, the “nucleus” would be smaller than in the case of

    参考来源

    - 位置:《Natural Language Processing in Action, Second Edition》第10章《真实世界中的大语言模型》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Natural Language Processing in Action, Second Edition/_epub-src/OEBPS/Text/chapter-10.html) - 结论依据:摘录中关于核采样按累计概率达到阈值截断的定义、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:Let’s try to generate text using the nucleus sampling method (listing 10.7). In this method, instead of choosing among the k most likely words, the model looks at the smallest set of words whose cumulative probability is smaller than p with the top_p parameter. So if there are only a few candidates with large probabilities, the “nucleus” would be smaller than in the case of