🛡️ Sentinella: Lightweight Content Safety Guardian

🎯 Model Overview

Sentinella is a compact yet powerful content safety classifier designed specifically for Italian language moderation. This model serves as your efficient first line of defense against harmful content.

📊 Key Metrics

• Size: 32M parameters
• Accuracy: 93% on test set
• Max Input Length: 8,192 tokens
• Training Data: more than 100,000 balanced examples (harmful/safe)

🔧 Technical Specifications

Base Architecture

• Base Model: jinaai/jina-embeddings-v2-small-en
• Model Adaptation:
  • Enhanced with a custom classifier head using a two-layer architecture
  • Optimized dropout rate of 0.1 for regularization
  • CLS token pooling strategy for sequence representation
  • Implemented with cross-entropy loss for binary classification

Classification Details

• Output Labels:
  • NEGATIVE (0): Harmful content
  • POSITIVE (1): Safe content

💫 Key Features

• Lightweight: At just 32M parameters, Sentinella is designed for efficiency
• Long Context: Handles up to 8k tokens of input text
• High Performance: 93% accuracy in content safety classification
• Optimized Architecture: Custom classification head with dimensionality reduction for improved efficiency

🚀 Use Cases

• Content moderation for Italian text
• Safe content filtering
• Automated content screening
• Real-time text analysis

🎓 Training Details

• Training Dataset: more than 100,000 examples
  • Balanced distribution of safe and harmful content
  • Focused on Italian language text
• Training Strategy:
  • Fine-tuned embedding representation
  • Intermediate layer dimensionality reduction
  • ReLU activation for non-linearity
  • Optimized dropout for regularization

📈 Performance Considerations

• Optimized for real-time classification
• Low memory footprint
• Efficient inference time
• Suitable for both CPU and GPU deployment

📝 Citation

If you use Sentinella in your research or application, please cite this work as:

@model{sentinella,
  title={Sentinella: Lightweight Italian Content Safety Classifier},
  year={2024},
  publisher={[Michele Montebovi]},
  note={32M parameter content safety model}
}