🧠 NeuroGraph: A Self-Aware Cognitive Router

NeuroGraph is a 17.8M parameter Deep-Narrow transformer designed to output structured knowledge graphs with per-node confidence scores.

🏢 Designed and engineered by CedzLabs

🎯 The Vision: Cascading Inference

Large Language Models are expensive and slow. NeuroGraph is designed as the "fast brain" in a cascading inference system:

1. Analyze: NeuroGraph instantly parses the user query and builds an initial knowledge graph.
2. Enrich: Low-confidence nodes trigger MCP (Model Context Protocol) tool calls to fetch live context (e.g., reading files, querying logs).
3. Escalate: Very low-confidence nodes trigger smart escalation to larger models (GPT-4, Claude) with rich, pre-packaged context.

📊 Model Specifications

Property	Value
Parameters	17.8 Million
Architecture	Deep-Narrow (24 layers, dim=256, 8 heads)
Positional Encoding	Split-Half RoPE (llama.cpp compatible)
Vocabulary	8,192 (Custom BPE)
Training Data	35,668 examples (Code, Reasoning, Knowledge, Synthetic)
GGUF Size (Q4_K_M)	~12.1 MB

🚀 Quick Start with Ollama

# Create a Modelfile
echo 'FROM ./neurograph-q4_k_m.gguf
PARAMETER num_ctx 512
PARAMETER temperature 0.7
PARAMETER repeat_penalty 1.5
SYSTEM "You are a cognitive router. Output knowledge graphs with nodes, edges, and confidence scores."' > Modelfile

# Create and run
ollama create neurograph -f Modelfile
ollama run neurograph "Debug this Python error: IndexError: list index out of range"

📝 Output Format

The model outputs structured JSON knowledge graphs:

{
  "output_type": "debug_graph",
  "nodes": [
    {
      "id": "error",
      "type": "issue",
      "content": "Python IndexError: list index out of range",
      "confidence": 0.95,
      "source": "model_knowledge"
    },
    {
      "id": "context",
      "type": "code",
      "content": "<NEEDS_MCP:read_file>",
      "confidence": 0.4,
      "source": "pending_mcp",
      "mcp_tool": "read_file",
      "mcp_args": {"path": "target.py"}
    }
  ],
  "edges": [
    {"from": "error", "to": "context", "relation": "occurs_in"}
  ]
}

⚠️ Limitations & Intended Use

This is a proof-of-concept research model developed by CedzLabs to demonstrate the viability of the Deep-Narrow Cognitive Router architecture.

• At 17.8M parameters, it is not designed for complex, multi-step reasoning.
• It is designed to be the first stage of a pipeline, identifying what it knows, what it needs to look up, and what it needs to escalate.
• Output text may contain repetition; the primary value is the structural schema and confidence scoring.

📄 License

MIT License. Free for research and commercial use.

🙏 Acknowledgments

• Architecture and training pipeline designed by CedzLabs
• Built with PyTorch, llama.cpp, and Hugging Face datasets