知识卡片

普通读书笔记卡 · 6511.a · 生成模型:GPT-6

优势只比较同一题组内表现

理解等级:中级

A_i=(r_i-组均值)/(组标准差+ε) 表示相对本题其他候选的表现,正值高于均值、负值低于均值。它不是跨题的绝对正确性分数;不同难度题的组内奖励分布会影响信号尺度。

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

      We now move on from rewards to the so-called advantages, as shown in figure 6.15. While rewards tell us how well each individual rollout performed, advantages capture how a rollout performed relative to other rollouts generated for the same prompt.

    参考来源

    - 位置:《Build a Reasoning Model (From Scratch)》第06章《强化学习训练推理模型》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Build a Reasoning Model (From Scratch)/_epub-src/OEBPS/Text/chapter-6.html) - 结论依据:摘录中关于优势只比较同一题组内表现的定义、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:We now move on from rewards to the so-called advantages, as shown in figure 6.15. While rewards tell us how well each individual rollout performed, advantages capture how a rollout performed relative to other rollouts generated for the same prompt.