跳到主要内容
Documentation

What is MoM Model Family?

The MoM (Mixture of Models) Model Family is a curated collection of specialized, lightweight models designed for intelligent routing, content safety, and semantic understanding. These models power the core capabilities of Semantic Router, enabling fast, accurate, and privacy-preserving AI operations.

Docs indexEdit page
版本：最新版

What is MoM Model Family?

The MoM (Mixture of Models) Model Family is a curated collection of specialized, lightweight models designed for intelligent routing, content safety, and semantic understanding. These models power the core capabilities of Semantic Router, enabling fast, accurate, and privacy-preserving AI operations.

Overview​

The MoM family consists of purpose-built models that handle specific tasks in the routing pipeline:

  • Classification Models: Domain detection, PII identification, jailbreak detection
  • Embedding Models: Semantic similarity, caching, retrieval
  • Safety Models: Hallucination detection, content moderation
  • Feedback Models: User intent understanding, conversation analysis

All MoM models are:

  • Lightweight: 33M-600M parameters for fast inference
  • Specialized: Fine-tuned for specific routing tasks
  • Efficient: Many use LoRA adapters for minimal memory footprint
  • Open Source: Available on HuggingFace for transparency and customization

Model Categories​

1. Classification Models​

Domain/Intent Classifier​

  • Model ID: models/mmbert32k-intent-classifier-merged
  • HuggingFace: llm-semantic-router/mmbert32k-intent-classifier-merged
  • Purpose: Classify user queries into 14 MMLU categories (math, science, history, etc.)
  • Architecture: mmBERT-32K merged classifier (307M)
  • Use Case: Route queries to domain-specific models or experts

PII Detector​

  • Model ID: models/mmbert32k-pii-detector-merged
  • HuggingFace: llm-semantic-router/mmbert32k-pii-detector-merged
  • Purpose: Detect 17 PII entity types across 35 BIO labels
  • Architecture: mmBERT-32K merged token classifier (307M)
  • Use Case: Privacy protection, compliance, data masking

Jailbreak Detector​

  • Model ID: models/mmbert32k-jailbreak-detector-merged
  • HuggingFace: llm-semantic-router/mmbert32k-jailbreak-detector-merged
  • Purpose: Detect prompt injection and jailbreak attempts
  • Architecture: mmBERT-32K merged classifier (307M)
  • Use Case: Content safety, prompt security

Feedback Detector​

  • Model ID: models/mmbert32k-feedback-detector-merged
  • HuggingFace: llm-semantic-router/mmbert32k-feedback-detector-merged
  • Purpose: Classify user feedback into 4 types (satisfied, need clarification, wrong answer, want different)
  • Architecture: mmBERT-32K merged classifier (307M)
  • Use Case: Adaptive routing, conversation improvement

2. Embedding Models​

Embedding Pro (High Quality)​

  • Model ID: models/mom-embedding-pro
  • HuggingFace: Qwen/Qwen3-Embedding-0.6B
  • Purpose: High-quality embeddings with 32K context support
  • Architecture: Qwen3 (600M parameters)
  • Embedding Dimension: 1024
  • Use Case: Long-context semantic search, high-accuracy caching

Embedding Flash (Balanced)​

  • Model ID: models/mom-embedding-flash
  • HuggingFace: google/embeddinggemma-300m
  • Purpose: Fast embeddings with Matryoshka support
  • Architecture: Gemma (300M parameters)
  • Embedding Dimension: 768 (supports 512/256/128 via Matryoshka)
  • Use Case: Balanced speed/quality, multilingual support

Embedding Ultra (Default)​

  • Model ID: models/mom-embedding-ultra
  • HuggingFace: llm-semantic-router/mmbert-embed-32k-2d-matryoshka
  • Purpose: Long-context multilingual semantic similarity with 2D Matryoshka support
  • Architecture: mmBERT 2D Matryoshka (307M parameters)
  • Embedding Dimension: 768 (supports lower dimensions via Matryoshka)
  • Use Case: Default semantic caching, retrieval, and tools similarity

3. Hallucination Detection Models​

Halugate Sentinel​

  • Model ID: models/mom-halugate-sentinel
  • HuggingFace: LLM-Semantic-Router/halugate-sentinel
  • Purpose: First-stage hallucination screening
  • Architecture: BERT-base (110M)
  • Use Case: Fast hallucination detection, pre-filtering

Halugate Detector​

  • Model ID: models/mom-halugate-detector
  • HuggingFace: KRLabsOrg/lettucedect-base-modernbert-en-v1
  • Purpose: Accurate hallucination verification
  • Architecture: ModernBERT-base (149M)
  • Context Length: 8192 tokens
  • Use Case: Factual accuracy verification, grounding check

Halugate Explainer​

  • Model ID: models/mom-halugate-explainer
  • HuggingFace: tasksource/ModernBERT-base-nli
  • Purpose: Explain hallucination reasoning via NLI
  • Architecture: ModernBERT-base (149M)
  • Classes: 3 (entailment/neutral/contradiction)
  • Use Case: Explainable AI, hallucination analysis

Model Selection Guide​

By Use Case​

Use CaseRecommended ModelWhy
Domain routingmmbert32k-intent-classifier-merged14 MMLU categories, 32K context
Privacy protectionmmbert32k-pii-detector-merged17 entity types, 35 BIO labels, 32K context
Content safetymmbert32k-jailbreak-detector-mergedPrompt injection detection with merged mmBERT
Semantic cachingmom-embedding-ultraDefault 32K multilingual embeddings
Long-context searchmom-embedding-pro32K context, 1024-dim
Hallucination checkmom-halugate-detectorModernBERT, 8K context
User feedbackmmbert32k-feedback-detector-merged4 feedback types, merged mmBERT

By Performance Requirements​

RequirementModel TierExamples
Ultra-fast (<10ms)Lightmom-embedding-flash, mmbert32k-jailbreak-detector-merged
Balanced (10-50ms)Defaultmom-embedding-ultra, mmbert32k-intent-classifier-merged
High-quality (50-200ms)Promom-embedding-pro, mom-halugate-detector

Configuration​

Using MoM Models in Router​

MoM models are configured through the canonical global.model_catalog block, with module-level settings living under global.model_catalog.modules:

global:
  model_catalog:
    system:
      domain_classifier: "models/mmbert32k-intent-classifier-merged"
      pii_classifier: "models/mmbert32k-pii-detector-merged"
      prompt_guard: "models/mmbert32k-jailbreak-detector-merged"
    modules:
      classifier:
        domain:
          model_ref: "domain_classifier"
          threshold: 0.6
          use_cpu: true
        pii:
          model_ref: "pii_classifier"
          threshold: 0.9
          use_cpu: true
      prompt_guard:
        model_ref: "prompt_guard"
        threshold: 0.7
        use_cpu: true

Custom System Bindings​

Override the built-in system-model bindings in your config.yaml:

global:
  model_catalog:
    system:
      domain_classifier: "models/your-domain-classifier"
      pii_classifier: "models/your-pii-classifier"
      prompt_guard: "models/your-prompt-guard"

Model Architecture​

LoRA-Based Models​

Many MoM models use LoRA (Low-Rank Adaptation) for efficiency:

  • Base Model: BERT-base-uncased (110M parameters)
  • LoRA Adapters: <1M parameters per task
  • Memory Footprint: ~440MB base + ~4MB per adapter
  • Inference Speed: Same as base model (~10-20ms on CPU)

ModernBERT Models​

Newer models use ModernBERT for better performance:

  • Architecture: ModernBERT-base (149M parameters)
  • Context Length: 8192 tokens (vs 512 for BERT)
  • Performance: Better accuracy on long-context tasks
  • Use Cases: Hallucination detection, feedback classification

Next Steps​

  • Signal-Driven Decisions - Learn how MoM models power routing decisions
  • Domain - Use mmbert32k-intent-classifier-merged for routing
  • PII - Configure mmbert32k-pii-detector-merged
  • RAG - Use MoM embedding models for route-local retrieval