← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8700.c · 生成模型:GPT-6

负例阈值方向取决于分数定义

理解等级:高级

相似度越高越相关时,负例一般应比正例低足够间隔,例如s_neg≤s_pos−m。若用距离则方向相反;原书“正例分数加margin”表述不能无条件套用,固定0.7或排名区间也不是消除假负例的保证。

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

      in a document corpus using an embedding model that are not already determined as positive examples. To ensure that the extracted examples are not false negatives, i.e., they are more relevant or just as relevant as the positive example, we can use a relevance score threshold (retrieve only examples with cosine similarity below 0.7) or a top-k range (only retrieve examples between top-30 and top-50).

    参考来源

    - 位置:《Designing Large Language Model Applications》第11章《表示学习与嵌入》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch11.html) - 结论依据:摘录中关于负例阈值方向取决于分数定义的定义、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:in a document corpus using an embedding model that are not already determined as positive examples. To ensure that the extracted examples are not false negatives, i.e., they are more relevant or just as relevant as the positive example, we can use a relevance score threshold (retrieve only examples with cosine similarity below 0.7) or a top-k range (only retrieve examples between top-30 and top-50).