知识卡片
跨度恢复用哨兵标识缺失块
T5式跨度破坏把连续缺失片段用各自哨兵替代,输出按哨兵顺序生成被删内容。哨兵用于对齐缺口与恢复段,目标还包含边界符号;不能把缺失token占比直接当全部编码解码计算成本的比例。
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
The T5 model creators used a modification of the original MLM objective. In this variant, 15% of tokens are randomly chosen to be removed from a sequence. Consecutive dropped-out tokens are replaced by a single unique special token called the sentinel token. The model is then asked to predict and generate the dropped tokens, delineated by the sentinel 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:摘录中关于跨度恢复用哨兵标识缺失块的定义、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:The T5 model creators used a modification of the original MLM objective. In this variant, 15% of tokens are randomly chosen to be removed from a sequence. Consecutive dropped-out tokens are replaced by a single unique special token called the sentinel token. The model is then asked to predict and generate the dropped tokens, delineated by the sentinel tokens.