← 返回刚才阅读的位置

知识卡片

结构图卡 · 6508.a · 生成模型:GPT-6

一致性投票依赖答案归并

理解等级:中级

结构图:

flowchart LR
 P[同一题目] --> C[生成多个候选]
 C --> X[抽取终值]
 X --> N[按协议归并等价答案]
 N --> V[统计票数]
 V --> R[唯一最高票或平票处理]

投票聚合的是答案,不是措辞各异的整段解释。书中直接 Counter 原始抽取字符串,等价分数、空结果等可能被错误拆分或归并,应先定义答案规范和无效票策略。

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

      The idea behind self-consistency sampling was formally introduced in the Google Research paper Self-Consistency Improves Chain-of-Thought Reasoning in Language Models (https://arxiv.org/abs/2203.11171). However, despite the fancy name, it's essentially a form of simple majority voting, where we use temperature scaling and top-p filtering to generate multiple answers and then select the most frequent one, as shown in figure 4.17.

    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:摘录中关于一致性投票依赖答案归并的定义、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:The idea behind self-consistency sampling was formally introduced in the Google Research paper Self-Consistency Improves Chain-of-Thought Reasoning in Language Models (https://arxiv.org/abs/2203.11171). However, despite the fancy name, it's essentially a form of simple majority voting, where we use temperature scaling and top-p filtering to generate multiple answers and then select the most frequent one, as shown in figure 4.17.