moderat - Dual-Mode Content Moderation + PII Filter

A text classification model for content moderation with age-appropriate filtering and PII detection.

Features

• Dual-mode filtering: <13 (strict) vs 13+ (laxed)
• 6 content categories: Safe, Harassment, Swearing (reaction), Swearing (aggressive), Hate Speech, Spam
• PII Detection: Emails, phones, addresses, credit cards, SSN
• Unicode Deobfuscation: Detects circled letters (ⓕ), double-struck (ℂ), fullwidth, mathematical symbols
• Social Media Protection:
  • <13: Block all social media sharing
  • 13+: Allow, block only if grooming detected
• Grooming Detection: Keywords like "dm me", "don't tell parents", "our secret"

Quick Start

from pii_extension import CombinedModerationFilter

filter = CombinedModerationFilter("darwinkernelpanic/moderat")

# Content moderation
result = filter.check("damn that's crazy", age=15)
# -> ALLOWED (reaction swearing for 13+)

# PII blocking (all ages)
result = filter.check("My email is test@gmail.com", age=15)
# -> BLOCKED (PII detected)

# Social media (13+ allowed)
result = filter.check("Follow me on instagram @user", age=15)
# -> ALLOWED

# Grooming detection
result = filter.check("DM me privately, don't tell parents", age=14)
# -> BLOCKED (grooming detected)

Unicode Deobfuscation

Automatically detects and normalizes unicode bypass attempts:

Technique	Example	Normalized
Circled letters	ⓕⓤⓒⓚ	fuck
Double-struck	ℂℍ	CH
Fullwidth	Ｆ	F
Mathematical	𝐟	f

All obfuscated text is normalized before moderation checks.

Social Media Rules

Age	Plain Share	With Grooming Context
<13	❌ Blocked	❌ Blocked
13+	✅ Allowed	❌ Blocked

Grooming keywords: "dm me", "don't tell", "secret", "send pics", "meet up", etc.

Content Labels

Text	<13	13+
"damn that's crazy"	❌ Blocked	✅ Allowed
"shit that sucks"	❌ Blocked	✅ Allowed
"you're trash"	❌ Blocked	❌ Blocked
"kill yourself"	❌ Blocked	❌ Blocked

Model Details

• Algorithm: Multinomial Naive Bayes + TF-IDF + Regex PII
• Content accuracy: 77%
• PII detection: Regex-based (fast, no ML)
• Features: 10,000 max, 1-3 ngrams

Files

• moderation_model.pkl - Content moderation model
• pii_extension.py - PII + grooming detection
• inference.py - Basic inference
• moderat_speed_test.ipynb - Colab notebook

Colab

Test it: Open in Colab

Speed

• Single inference: ~2-5ms
• With PII check: ~3-7ms
• Throughput: ~300-500 texts/sec