h3rb3rn / moe-expert-graphrag-4b

🕸️ MoE Sovereign GraphRAG Expert 4B ( moe-expert-graphrag-4b )

Multi-Hop Knowledge Graph Traversal, Cypher Query Generation & Entity Resolution

📌 Executive Summary & Architectural Role

moe-expert-graphrag-4b is a specialized 4-billion parameter Small Language Model (SLM), LoRA fine-tuned on the LUMI-G Supercomputer (8× AMD Instinct™ MI250X GCDs (4× physical modules, 64GB HBM2e per GCD)).

Within the MoE Sovereign compound AI architecture, this model serves as the Knowledge Graph Navigation & GraphRAG Retrieval Specialist . It does not attempt to memorize static enterprise facts within its weights. Instead, it is trained on the operational mechanics of structured knowledge retrieval: compiling natural language questions into multi-hop Cypher queries, resolving fuzzy entity identifiers, performing hybrid vector-graph fusion, and extracting semantic triples (Subject, Predicate, Object) from unstructured documents.

🎯 Functional Scope & Capabilities

1. Multi-Hop Cypher Query Generation: Formulates syntactically valid openCypher / Neo4j 5.x graph queries ( MATCH , WHERE , OPTIONAL MATCH , WITH , RETURN ).
2. Entity Resolution & Linking: Maps ambiguous colloquial mentions in user prompts to canonical entity nodes in the enterprise knowledge graph.
3. Structured Triplet Extraction: Deconstructs unstructured text into validated knowledge graph triples with provenance metadata.
4. Vector + Graph Hybrid Fusion: Synthesizes dense semantic embeddings with explicit relational graph topologies.

🎯 Training Objectives & Intended Behavioral Specialization

📊 Empirical Evaluation (Held-Out Benchmark Suite)

ℹ️ Evaluation Status: Evaluated on held-out validation splits ($N=1,000$, zero training contamination). Full cross-architecture ablation suites across Compound AI vs. Monolithic LLMs are undergoing active execution in the Sovereign Scientific Benchmark Suite v1.

Evaluated on a held-out benchmark suite of 1,000 graph retrieval and Cypher generation tasks verified against a live Neo4j 5.25 graph instance with zero training contamination:

Note: Evaluated at temperature=0.0 across 3 independent seeds. Executable queries were executed directly against a multi-tenant enterprise ontology graph.

🏋️ Training Setup

+-----------------------------------------------------------------------------------+
|                          LUMI-G LORA FINE-TUNING PIPELINE                         |
|  [ Student: Qwen3.5-4B Hybrid Linear Attention + Mamba Base ]                     |
|                       |                                                           |
|                       v  (LoRA r=16, alpha=32, target_modules: q/k/v/o/gate/up/down)|
|  [ Output: final_adapter -> CPU-BF16 Merge -> GGUF Q4_K_M & Q8_0 ]                |
+-----------------------------------------------------------------------------------+

Hyperparameters:

• Compute Cluster: LUMI-G (8× AMD Instinct MI250X 128GB GPUs)
• Base Architecture: Qwen3.5-4B (Hybrid Linear Attention + Mamba in BF16)
• Epochs: 3.0
• Effective Batch Size: 128 (Micro-batch 4 × 8 GPUs × Gradient Accumulation 4)
• Learning Rate: $1.5 \times 10^{-5}$ with Cosine Decay and Warmup
• LoRA Configuration: $r=16$, $\alpha=32$, Dropout $0.05$, Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj

⚠️ Known Limitations & Failure Modes

1. Schema Visibility Requirement: The model relies on the active schema/ontology definition being supplied in the context; without schema hints, complex domain-specific relationship types cannot be deduced.
2. Unbounded Graph Cartesian Products: While the model is trained to avoid cartesian products ( MATCH (a), (b) without predicates), complex graph aggregations should be safeguarded by DB query timeouts.
3. Graph Topology Size: Queries returning more than 10,000 nodes are best streamed through cursor pagination rather than loaded in a single context window.

💻 Quickstart Guide (Ollama & Llama.cpp)

1. Ollama Modelfile

FROM ./moe-expert-graphrag-4b-Q4_K_M.gguf
PARAMETER num_ctx 262144
PARAMETER temperature 0.0
TEMPLATE """{{ if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}{{ if .Prompt }}<|im_start|>user
{{ .Prompt }}<|im_end|>
{{ end }}<|im_start|>assistant
{{ .Response }}<|im_end|>"""

2. Python Inference

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "h3rb3rn/moe-expert-graphrag-4b"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)

prompt = "<|im_start|>user\nGenerate a parameterized Neo4j Cypher query to find all microservices that depend on Kafka cluster 'kafka-prod-01' up to 3 hops.<|im_end|>\n<|im_start|>assistant\n"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.0)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

📑 Citation

@misc{moe_sovereign_2026_graphrag4b,
  author = {Horn, Philipp and MoE Sovereign Core AI Team},
  title = {MoE Sovereign GraphRAG Expert 4B: Knowledge Graph Traversal & Cypher SLM},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/h3rb3rn/moe-expert-graphrag-4b}},
  note = {Trained on the EuroHPC LUMI-G Supercomputer}
}