知识卡片

普通读书笔记卡 · 6730.c · 生成模型:GPT-6

重复记录与过采样不能无条件等同

理解等级:中级

重复行可能来自原始采样或处理错误,只有按明确方案增加少数类样本才是有意过采样。去重会改变频率,保留又会改变损失权重;划分前还需按共享对象隔离,避免重复同时进入训练与验证。

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 addition to the overlap across nationalities, the PyTorch tutorial dataset contained many duplicated names within nationalities. More than 94% of the Arabic names were duplicates, some of which are shown in listing 8.5. Other nationalities and languages, such as English, Korean, and Scottish, appear to have been deduplicated. Duplicates in your training set make your model fit more closely to common names than to

    参考来源

    - 位置:《Natural Language Processing in Action, Second Edition》第08章《循环网络与长短期记忆》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Natural Language Processing in Action, Second Edition/_epub-src/OEBPS/Text/chapter-8.html) - 结论依据:摘录中关于重复记录与过采样不能无条件等同的定义、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 addition to the overlap across nationalities, the PyTorch tutorial dataset contained many duplicated names within nationalities. More than 94% of the Arabic names were duplicates, some of which are shown in listing 8.5. Other nationalities and languages, such as English, Korean, and Scottish, appear to have been deduplicated. Duplicates in your training set make your model fit more closely to common names than to