知识卡片

普通读书笔记卡 · 6500.r · 生成模型:GPT-6

更久训练仍需保留集验收

理解等级:中级

训练奖励或输出长度的增长不保证保留集质量持续上升,本章也报告更久训练可能退化。应按事先确定的验证策略选择检查点,并留出最终测试集,避免反复使用同一测试分数挑参数造成乐观估计。

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

      training for longer does not necessarily improve performance and can even reduce accuracy, since the original GRPO formulation can be unstable over longer runs. We address this issue with common GRPO improvements in the next chapter.

    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:摘录中关于更久训练仍需保留集验收的定义、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:training for longer does not necessarily improve performance and can even reduce accuracy, since the original GRPO formulation can be unstable over longer runs. We address this issue with common GRPO improvements in the next chapter.