知识卡片
预处理顺序依赖实际输入
数据管线包含抽取、清洗、质量筛选、去重、隐私处理与去污染,但先后取决于各步骤所需表示。应明确每步输入输出与删除原因,避免后续统计把已过滤数据当完整来源;大规模处理还需考虑每步成本。
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
Figure 2-3 shows the typical preprocessing steps used to generate a pre-training dataset. The ordering of steps is not fixed, but there are dependencies between some of the steps. Figure 2-3. Data collection and preprocessing pipeline Let’s go through these steps in detail. Data Filtering and Cleaning A majority of text extracted from HTML files is gibberish, like menu text from websites, boilerplate text, and
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:摘录中关于预处理顺序依赖实际输入的定义、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:Figure 2-3 shows the typical preprocessing steps used to generate a pre-training dataset. The ordering of steps is not fixed, but there are dependencies between some of the steps. Figure 2-3. Data collection and preprocessing pipeline Let’s go through these steps in detail. Data Filtering and Cleaning A majority of text extracted from HTML files is gibberish, like menu text from websites, boilerplate text, and