| Property | Value |
|---|---|
| Author | p1atdev |
| Model Type | Text Classification |
| Parameters | 30M |
| Language | All Languages |
| License | CC-BY-NC-SA-4.0 |
| Base Model | sbintuitions/modernbert-ja-30m |
What is saikyou-shield-30m?
Saikyou Shield 30M is an April Fools' 2025 text classification model designed with a unique approach: it classifies every input prompt as "dangerous". Built on ModernBERT architecture, this lightweight 30M parameter model achieves a perfect 100% detection rate by considering all inputs as potentially harmful.
Implementation Details
The model is implemented using the ModernBERT architecture and can be easily deployed using the Transformers library. It was trained on 1,000 samples from the japanese_alpaca_data dataset, with all samples deliberately labeled as "dangerous". Training was performed with a learning rate of 1e-2 for binary classification.
- Extremely lightweight (30M parameters)
- 100% detection rate for all inputs
- Compatible with all languages
- Simple integration with Transformers pipeline
Core Capabilities
- Classifies any input text as dangerous with 100% confidence
- Reduces API costs by blocking all inputs
- Runs efficiently on any device, including MacBook Air M3
- Perfect for demonstrations of over-aggressive content filtering
Frequently Asked Questions
Q: What makes this model unique?
This model takes content filtering to its logical extreme by classifying everything as dangerous, achieving perfect recall at the cost of precision. It's a humorous take on content moderation systems.
Q: What are the recommended use cases?
While created as an April Fools' joke, the model demonstrates the concept of over-aggressive content filtering. It's not intended for actual production use, as it cannot identify genuinely safe content.