知识卡片

普通读书笔记卡 · 6512.b · 生成模型:GPT-6

损失数值不代表任务准确率

理解等级:中级

策略损失可正、可负或接近零,随优势分布、长度和模型概率变化,不能像统一误差率一样跨步直接解释。监控应同时看平均奖励、格式/终值分量、组内奖励方差、生成长度与保留集表现。

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

      loss values themselves should not be over-interpreted. Steps with reward_avg=0.000 produce a near-zero loss because all rollouts receive the same reward, resulting in vanishing group-relative advantages and little to no gradient signal. Larger loss magnitudes, such as at step 3, simply reflect bigger relative differences between rollouts and are typical for GRPO-style objectives, especially at the beginning of training.

    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:摘录中关于损失数值不代表任务准确率的定义、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:loss values themselves should not be over-interpreted. Steps with reward_avg=0.000 produce a near-zero loss because all rollouts receive the same reward, resulting in vanishing group-relative advantages and little to no gradient signal. Larger loss magnitudes, such as at step 3, simply reflect bigger relative differences between rollouts and are typical for GRPO-style objectives, especially at the beginning of training.