[Feature Request] Improve Toxicity Scanner for Russian Language and Custom Use Cases #284
Description
Current Situation
I'm working on an educational project with the toxicity guard system using llm-guard. I've encountered several issues that need improvement while working with other languages like russian for example :
1. Russian Language Support Issues
- The current toxicity scanner doesn't properly detect Russian-specific toxic content
- Russian chemical terms (like "фозген" - phosgene) are not recognized as potentially dangerous
- The scanner seems to be primarily trained on English content
2. Custom Use Case Requirements
I need to add specific toxicity indicators for my particular use case:
- Chemical substance detection (especially dangerous chemicals)
- Scientific terminology that could be misused
- Russian-specific toxic language patterns
- Context-aware toxicity detection
3. Current Scanner Behavior
From my testing, the scanner:
- ✅ Detects basic English toxicity patterns
- ❌ Misses Russian language toxic content
- ❌ Doesn't recognize domain-specific dangerous terms
- ❌ Has high false positives for Russian text
What I Need
2. Custom Toxicity Indicators
- Ability to add custom toxic patterns
- Domain-specific toxicity detection (chemical, medical, etc.)
- Configurable toxicity thresholds per language
Technical Details
Current Setup:
- Using llm-guard with GigaChat API
- Python 3.12 on macOS
- All dependencies installed correctly