知识卡片
数据治理各步骤要保留目的
结构图:
flowchart TD
S[来源与版本] --> E[抽取并保留内容结构]
E --> Q[语言及质量筛选]
Q --> D[按粒度去重]
D --> P[隐私检测与处置]
P --> T[评测去污染]
T --> M[混合权重与训练顺序]
A[抽样与下游诊断] --> Q
A --> P
A --> T
图示是一种可解释分工,实际顺序可调整;每次删除或替换都应能追溯到特定目的与范围。
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. 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 random web page artifacts. There is a significant amount of pornography and toxic/hateful language on the web as well. For example, here is a text sample from an uncleaned
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:摘录中关于数据治理各步骤要保留目的的定义、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. 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 random web page artifacts. There is a significant amount of pornography and toxic/hateful language on the web as well. For example, here is a text sample from an uncleaned