← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 6512.d · 生成模型:GPT-6

训练和评估长度分别报告

理解等级:中级

本章训练每条 rollout 的最大长度与评估最大长度不同,评估平均回答也可长于训练预算。表格应分列训练设置和评估协议,不能把评估超出训练上限误认为日志矛盾,也不能在不同预算下直接归因模型改进。

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

      maximum token limit of 2048, allowing the LLM to generate longer responses during evaluation. The "Average tokens" column averages the response length during the evaluation on the MATH-500 dataset. What we can see is that, compared to the reference reasoning model (row 2), our GRPO models generate shorter responses on average, as expected due to the token-length restriction during 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:摘录中关于训练和评估长度分别报告的定义、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:maximum token limit of 2048, allowing the LLM to generate longer responses during evaluation. The "Average tokens" column averages the response length during the evaluation on the MATH-500 dataset. What we can see is that, compared to the reference reasoning model (row 2), our GRPO models generate shorter responses on average, as expected due to the token-length restriction during training.