Skip to main content

title-cleaner.py

Path: build/title-cleaner.py | Language: Python | Lines: ~882

Removes cruft from webpage titles extracted from HTML <title> tags using GPT-5-mini language model.

Overview​

When scraping HTML pages, the <title> field often contains unwanted content: site names, domain suffixes, error messages, boilerplate text, or garbled Unicode. Manually writing rules to clean >20,000 URLs on gwern.net would be unmanageable due to the extreme long-tail of edge cases.

This script uses GPT-5-mini to intelligently clean titles based on "I know it when I see it" heuristics. It removes spam, cruft, and boilerplate while preserving meaningful content. The script can also convert inline Markdown (like *emphasis*) to HTML (<em>emphasis</em>).

The cleaning process is conservative: when uncertain, it returns the original title unchanged. Titles deemed completely useless (error pages, generic "Index", empty strings, site-wide constants) are replaced with empty strings for manual review.

Note: This script does not add italics for titles of works (books, movies, etc.) even when obviously appropriate, as English italicization rules are complex and context-dependent. That task is handled separately by italicizer.py. Both scripts accept stdin or arguments and can be chained: echo "Title" | title-cleaner.py | italicizer.py.

Key Functions​

  • Main processing: Reads input from stdin or command-line argument, sends to GPT-5-mini
  • Extensive prompt examples: 600+ few-shot examples covering common title problems
  • Output format: Returns cleaned title or empty string (never adds commentary)
  • Markdown-to-HTML conversion: Converts *foo* → <em>foo</em> (inline markdown only)

Common title problems handled:

  • Site name suffixes: "Article Title - The New York Times" → "Article Title"
  • Error messages: "404", "502 Bad Gateway", "Page not found" → ""
  • Boilerplate: "Gwern.net | Article" → "Article"
  • Garbled Unicode: "miserâs heirs" → "miser's heirs"
  • Useless constants: "index", "articles", "Reddit", "Home" → ""
  • Security checkpoints: "Vercel Security Checkpoint" → ""
  • Generic domains: "Notion – The all-in-one workspace..." → ""

Command Line Usage​

# From stdin
echo "If I Sleep for an Hour, 30 People Will Die - The New York Times" | OPENAI_API_KEY="sk-XXX" python title-cleaner.py
# Output: If I Sleep for an Hour, 30 People Will Die

# From command-line argument
OPENAI_API_KEY="sk-XXX" python title-cleaner.py "404"
# Output: (empty string)

# Chain with italicizer
echo "Review of The Hobbit - Fantasy Books" | title-cleaner.py | italicizer.py
# Output: Review of <em>The Hobbit</em>

# Handle Unicode corruption
echo "After 92 years, millionaire miserâs heirs finally split $100M - TODAY.com" | OPENAI_API_KEY="sk-XXX" python title-cleaner.py
# Output: After 92 years, millionaire miser's heirs finally split $100M

# Handle inline Markdown
echo "Anton Seder's *The Animal in Decorative Art* (1896)" | OPENAI_API_KEY="sk-XXX" python title-cleaner.py
# Output: Anton Seder's <em>The Animal in Decorative Art</em> (1896)

Requirements:

  • Python 3
  • openai Python package
  • Valid OpenAI API key in $OPENAI_API_KEY environment variable

Edge cases:

  • Preserves important context: "It's (Still) Really Hard for Robots..."
  • Removes author bylines: "by Robin Sloan The Message" → main title only
  • Handles special characters: &amp; preserved, \226\128\153 decoded
  • Recognizes error pages: "Page not found : Stanford University" → ""
  • Domain-only titles: "Flashback Forum", "GoLocalPDX", "Notion" → ""

See Also​