← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 6636.b · 生成模型:GPT-6

掩码目标只在选定位置计分

理解等级:高级

MLM选择部分输入位置作为恢复目标,可利用左右上下文。原始BERT不是把所选15%全部替换MASK,而是在这些位置按80%MASK、10%随机词、10%不变处理;选择目标位置与如何扰动它们是两层规则。

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 the MLM setting, rather than predict the next token in a sequence, we ask the model to predict masked tokens within the sequence. In the most basic form of MLM implemented in the BERT model, 15% of tokens are randomly chosen to be masked and are replaced with a special mask token, and the language model is asked to predict the original tokens.

    参考来源

    - 位置:《Designing Large Language Model Applications》第04章《架构与学习目标》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch04.html) - 结论依据:摘录中关于掩码目标只在选定位置计分的定义、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 the MLM setting, rather than predict the next token in a sequence, we ask the model to predict masked tokens within the sequence. In the most basic form of MLM implemented in the BERT model, 15% of tokens are randomly chosen to be masked and are replaced with a special mask token, and the language model is asked to predict the original tokens.