知识卡片
零梯度不保证参数不动
当前策略梯度为零不等于 optimizer.step 不会改变权重:优化器动量历史或 AdamW 的解耦权重衰减仍可能作用。判断是否跳过更新,应按实际优化器状态与策略定义,不能仅根据标量 loss 为零下结论。补充依据:AdamW 更新公式。
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
and 'text' entries in the 'samples' field, the model answers incorrectly. As discussed earlier in section 6.4, a correct answer must include \boxed{6}. Since this condition is not met, the reward is zero, which in turn yields zero advantages and a zero loss. In this case, if we were training the model, the gradient would be zero and the model parameters would not be updated.
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:摘录中关于零梯度不保证参数不动的定义、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:and 'text' entries in the 'samples' field, the model answers incorrectly. As discussed earlier in section 6.4, a correct answer must include \boxed{6}. Since this condition is not met, the reward is zero, which in turn yields zero advantages and a zero loss. In this case, if we were training the model, the gradient would be zero and the model parameters would not be updated.