知识卡片
训练奖励会继承判分漏洞
同一验证器既可用于离线评估,也可变成强化学习奖励来源。若抽取或判等存在捷径,优化过程可能放大这些缺口。因此在把布尔正确性转成奖励前,需独立检查其语义范围、异常策略和可被利用的格式行为。
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
- · 评估推理模型
原书出处 · 1 处
- · 查看原书依据
While our immediate focus is evaluation, verifiers will reappear later in this book. They not only serve as a way to measure performance but also provide the feedback signal used in reinforcement learning methods for training reasoning models, which we will explore in chapter 6.
参考来源
- 位置:《Build a Reasoning Model (From Scratch)》第03章《评估推理模型》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Build a Reasoning Model (From Scratch)/_epub-src/OEBPS/Text/chapter-3.html)
- 结论依据:摘录中关于训练奖励会继承判分漏洞的定义、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:While our immediate focus is evaluation, verifiers will reappear later in this book. They not only serve as a way to measure performance but also provide the feedback signal used in reinforcement learning methods for training reasoning models, which we will explore in chapter 6.