From a web page to a clean document Diagram

Extraction is not just stripping HTML tags

A naive extractor keeps the nav, cookie banner and footer
An overly aggressive extractor deletes code blocks, tables and formulas
Math or code pipelines need to keep MathJax, KaTeX, MathML

Normalization is a double-edged sword

Fixing encoding and dropping control characters is almost always right
Forced lowercasing or stripping diacritics can wreck Vietnamese and poetry
Whitespace is real syntax in Python and Markdown

Language ID is metadata, not a filter

It is used to control sampling ratios across languages
Mixed-language documents often get uncertain scores
A wrong language label flows straight into the later data-mixing step