知识卡片

普通读书笔记卡 · 8716.b · 生成模型:GPT-6

首词概率间隔不代表整答可靠

理解等级:高级

首token最高与次高概率之差可作为便宜路由特征,但可能主要反映开场格式和词汇选择。应实测它与完整答案质量的关联,不能仅因开头确定就认定多步推理或后续事实可靠。

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

      margin sampling, as proposed by Ramirez et al. In the margin sampling method, we generate the first token and use the difference in the probability of the most probable token and the second most probable token as the margin. The assumption is that the higher the margin, the more confident the model. If the margin is below a certain threshold, then the input is sent

    参考来源

    - 位置:《Designing Large Language Model Applications》第13章《设计模式与系统架构》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch13.html) - 结论依据:摘录中关于首词概率间隔不代表整答可靠的定义、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:margin sampling, as proposed by Ramirez et al. In the margin sampling method, we generate the first token and use the difference in the probability of the most probable token and the second most probable token as the margin. The assumption is that the higher the margin, the more confident the model. If the margin is below a certain threshold, then the input is sent