RAXE ML Models

ML models for RAXE Community Edition L2 threat detection.

Available Models

Model	Version	Size	Description
threat_classifier_gemma_mlp_v3	v0.4.0	~235MB	Gemma MLP 5-head classifier (INT8 quantized)

Current Model (v0.4.0)

The latest model features:

• Architecture: EmbeddingGemma-300M + 5-head MLP classifier
• Classification Heads:
  • Binary (threat/safe): 2 classes
  • Threat Family: 15 classes
  • Severity: 3 levels (none/moderate/severe)
  • Primary Technique: 35 classes
  • Harm Types: 10 classes (multilabel)
• Performance: TPR 91.2%, FPR 6.4%, F1 0.94
• Quantization: INT8 for efficient CPU inference

Installation

Models are automatically downloaded on first use:

pip install raxe
raxe scan "test prompt"  # Auto-downloads model on first use

Or download manually:

raxe models download

Model Files

The model package includes:

• model_int8.onnx - EmbeddingGemma-300M embeddings (INT8)
• classifier_is_threat_int8.onnx - Binary threat classifier
• classifier_threat_family_int8.onnx - 15-class family classifier
• classifier_severity_int8.onnx - 3-class severity classifier
• classifier_primary_technique_int8.onnx - 35-class technique classifier
• classifier_harm_types_int8.onnx - 10-class harm multilabel classifier
• tokenizer.json - Tokenizer configuration
• model_metadata.json - Model metadata and inference pipeline
• label_config.json - Label mappings for all heads
• pooling_config.json - Pooling configuration
• feature_scaler.pkl - Feature normalization

Performance

• Inference time: ~3-5ms per prompt (CPU)
• Memory footprint: ~400MB loaded
• Model download: ~235MB compressed

Changelog

v0.4.0 (2025-01-19)

• Updated to new classification schema
• Severity: 5 levels → 3 levels (none/moderate/severe)
• Threat Family: 9 classes → 15 classes
• Primary Technique: 22 classes → 35 classes
• Improved accuracy: TPR 91.2%, FPR 6.4%

v0.3.0 (2024-12-14)

• Initial Gemma MLP 5-head classifier
• TPR 90.4%, FPR 7.4%

License

These models are provided under the RAXE Community Edition license. See raxe-ai/raxe-ce for details.