知识卡片
前缀语言目标分开可见性与计分
PrefixLM允许前缀内部双向可见,后缀按自回归方式生成并计分。训练需同时明确遮罩与损失位置;双向读取已给定前缀不等于可以在预测后缀时偷看后缀未来。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
- · 架构与学习目标
原书出处 · 1 处
- · 查看原书依据
system like English, tokens do not attend to tokens to the right (future). In the prefix LM setting, a part of the text sequence, called the prefix, is allowed to attend to future tokens in the prefix. The prefix part is thus noncausal. For training prefix LMs, a random prefix length is sampled, and the loss is calculated over only the tokens in the suffix.
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:摘录中关于前缀语言目标分开可见性与计分的定义、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:system like English, tokens do not attend to tokens to the right (future). In the prefix LM setting, a part of the text sequence, called the prefix, is allowed to attend to future tokens in the prefix. The prefix part is thus noncausal. For training prefix LMs, a random prefix length is sampled, and the loss is calculated over only the tokens in the suffix.