知识卡片

普通读书笔记卡 · 6498.b.7 · 生成模型:GPT-6

格式约束会进入奖励目标

理解等级:中级

本章仅当终值正确且能从 boxed 格式抽取时才给 1,答案正确但未框住也得 0。训练因此同时优化格式与终值;监控奖励时宜分别记录两项,否则无法判断增长来自格式遵循还是解题能力。

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 rewards are computed using the math verifier from chapter 3 and are primarily based on answer correctness. By setting fallback=None in the reward_rlvr function in listing 6.6, we also include an implicit format constraint. For instance, an answer only receives a reward of 1.0 if it is both correct and expressed in the required \boxed{} format.

    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:摘录中关于格式约束会进入奖励目标的定义、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 rewards are computed using the math verifier from chapter 3 and are primarily based on answer correctness. By setting fallback=None in the reward_rlvr function in listing 6.6, we also include an implicit format constraint. For instance, an answer only receives a reward of 1.0 if it is both correct and expressed in the required \boxed{} format.