知识卡片
自监督从输入构造训练目标
大规模预训练可利用文本自身的后继或缺失片段作为目标,减少逐条人工标签的需要。没有人工标注不等于没有目标、没有选择偏差或数据无成本;采集、清洗与目标构造仍决定学到什么。
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
- · 预训练数据
原书出处 · 1 处
- · 查看原书依据
it has been shown that higher-capacity models are relatively more sample efficient, in general today’s language models are very sample inefficient, meaning they need tons of examples to learn a task. It is infeasible to create such a large supervised dataset with human annotations, hence the predominant means to pre-train language models is using self-supervised learning, where the target labels exist within your training inputs.
参考来源
- 位置:《Designing Large Language Model Applications》第02章《预训练数据》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch02.html)
- 结论依据:摘录中关于自监督从输入构造训练目标的定义、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:it has been shown that higher-capacity models are relatively more sample efficient, in general today’s language models are very sample inefficient, meaning they need tons of examples to learn a task. It is infeasible to create such a large supervised dataset with human annotations, hence the predominant means to pre-train language models is using self-supervised learning, where the target labels exist within your training inputs.