← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 6605.d · 生成模型:GPT-6

涌现结论依赖评估粒度

理解等级:高级

若评分只奖励完全正确,平滑改善的局部能力可能在最终分数上呈现突然跃升。分析所谓涌现时应同时看严格成功率与有意义的细粒度指标;指标影响可见曲线,不说明最终任务成功要求可以忽略。

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

      artifact of the evaluation metrics used to judge performance. This happens because many metrics do not assign partial credit and only reward fully solving the task, so model improvements might not be tracked. On the other hand, one could argue that for tasks like multi-step arithmetic, getting the answer partially right is just as useless as getting it completely wrong.

    参考来源

    - 位置:《Designing Large Language Model Applications》第01章《大语言模型入门》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch01.html) - 结论依据:摘录中关于涌现结论依赖评估粒度的定义、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:artifact of the evaluation metrics used to judge performance. This happens because many metrics do not assign partial credit and only reward fully solving the task, so model improvements might not be tracked. On the other hand, one could argue that for tasks like multi-step arithmetic, getting the answer partially right is just as useless as getting it completely wrong.