知识卡片
本章目标是教学策略梯度
本章代码使用负的“优势×序列对数概率”均值,省略完整 GRPO 的多项机制。原始 DeepSeekMath 式(3)包含逐 token 概率比、裁剪、长度归一化与 KL 项;因此本章不能被解释为只去掉 KL 的原始 GRPO。补充依据:DeepSeekMath 4.1.1。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
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 GRPO outline in figure 6.9 contains many technical components, and at first glance, it may look overwhelming. We will start by implementing a simplified version of GRPO that omits the KL loss term shown on the right side of the figure.
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:摘录中关于本章目标是教学策略梯度的定义、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 GRPO outline in figure 6.9 contains many technical components, and at first glance, it may look overwhelming. We will start by implementing a simplified version of GRPO that omits the KL loss term shown on the right side of the figure.