知识卡片
混合去噪覆盖不同预测形态
UL2式混合通过不同跨度、破坏比例和前后缀安排训练同一模型。常规跨度、顺序后缀和更高破坏率使模型接触不同信息缺失形态;它整合训练目标,不自动保证所有任务均优于专用模型。
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
UL2 mimics the effect of PLMs, MLMs, and PrefixLMs in a single paradigm called Mixture of Denoisers. The denoisers used are as follows: R-Denoiser This is similar to the T5 span corruption task. Spans between length 2–5 tokens are replaced by a single mask token. Figure 4-14 depicts the workings of the R-denoiser. Figure 4-14. UL2’s R-Denoiser S-Denoiser Similar to prefix LM, the text is
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:摘录中关于混合去噪覆盖不同预测形态的定义、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:UL2 mimics the effect of PLMs, MLMs, and PrefixLMs in a single paradigm called Mixture of Denoisers. The denoisers used are as follows: R-Denoiser This is similar to the T5 span corruption task. Spans between length 2–5 tokens are replaced by a single mask token. Figure 4-14 depicts the workings of the R-denoiser. Figure 4-14. UL2’s R-Denoiser S-Denoiser Similar to prefix LM, the text is