← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 6500.g.3 · 生成模型:GPT-6

日志重算指标保留分词配置

理解等级:中级

已存生成原文可离线重算长度或重判分,省去再次调用模型;但重编码长度需要固定 tokenizer 和模板设置,避免把新增控制标记计入原始输出。最可靠的生成长度来自当次记录的实际 token IDs 或计数,而非事后默认配置。

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

      Alternatively, the second option is to calculate the response lengths from the .jsonl files that were created when we ran the evaluate_math500_stream function in the main chapter. This way, we avoid having to rerun the evaluation.

    参考来源

    - 位置:《Build a Reasoning Model (From Scratch)》第08章《附录B 练习答案》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Build a Reasoning Model (From Scratch)/_epub-src/OEBPS/Text/appendix-b.html) - 结论依据:摘录中关于日志重算指标保留分词配置的定义、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:Alternatively, the second option is to calculate the response lengths from the .jsonl files that were created when we ran the evaluate_math500_stream function in the main chapter. This way, we avoid having to rerun the evaluation.