MoE Sovereign Knowledge-Graph & GraphRAG Retrieval Expert 4B (moe-expert-graphrag-4b)
Model Summary
moe-expert-graphrag-4b is a LoRA fine-tune of the text-decoder of Qwen3.5-4B, specialized as the graphrag domain expert within the MoE Sovereign compound-AI system.
You are a knowledge-graph navigation and GraphRAG retrieval specialist (moe-expert-graphrag-4b). Formulate syntactically correct multi-hop Cypher queries, resolve ambiguous entity mentions to canonical nodes, and extract structured knowledge triplets from unstructured text. Require the graph schema to be given in context -- never invent node or relationship types that weren't provided, and bound multi-hop traversals against runaway cartesian products.
Base Architecture
Qwen3.5-4B is a hybrid linear-attention / full-attention decoder (not a plain Transformer): 32 layers (8 full-attention, 24 linear/Mamba-style), hidden size 2,560, 248,320-token vocabulary, native 262,144-token context window.
Training Configuration
| Parameter | Value |
|---|---|
| Method | LoRA (rank 16, alpha 32, dropout 0.05), targeting q/k/v/o_proj + gate/up/down_proj |
| Trainable parameters | 21,233,664 of 4,226,984,960 (0.50%) |
| Epochs | 3 |
| Effective batch size | 128 (micro-batch 4 x 8 GPUs x grad-accum 4) |
| Learning rate | 1.5e-5 |
| Training sequence length | 4,096 tokens |
| Optimizer sharding | DeepSpeed ZeRO-2, bf16 |
| Compute | EuroHPC LUMI-G, 8x AMD Instinct MI250X GCDs, ROCm |
| Training examples | 5,277 curated instruction/response pairs |
Observed Training Trajectory
Training loss over the run (representative logged steps): 1.913 -> 0.9435 -> 0.8844. Smooth, monotonic decline consistent with genuine generalization, not memorization.
Prompt Format
ChatML:
<|im_start|>system
{system_prompt}<|im_end|>
<|im_start|>user
{user_message}<|im_end|>
<|im_start|>assistant
{response}<|im_end|>
System Prompt
You are a knowledge-graph navigation and GraphRAG retrieval specialist (moe-expert-graphrag-4b). Formulate syntactically correct multi-hop Cypher queries, resolve ambiguous entity mentions to canonical nodes, and extract structured knowledge triplets from unstructured text. Require the graph schema to be given in context -- never invent node or relationship types that weren't provided, and bound multi-hop traversals against runaway cartesian products.
Available Formats
| File | Notes |
|---|---|
moe-expert-graphrag-4b-Q4_K_M.gguf |
Recommended for single/multi-GPU deployment |
moe-expert-graphrag-4b-Q8_0.gguf |
Higher-fidelity reference quantization |
Hardware Guidance
Native 262,144-token context usable in full on multi-GPU pools with q4_0-quantized KV-cache and Flash Attention (Ampere/Turing+). On single 8GB GPUs cap num_ctx to 32,768 and use f16 KV-cache (Maxwell-generation GPUs lack Flash Attention support).
Ollama Modelfile
FROM ./moe-expert-graphrag-4b-Q4_K_M.gguf
SYSTEM """You are a knowledge-graph navigation and GraphRAG retrieval specialist (moe-expert-graphrag-4b). Formulate syntactically correct multi-hop Cypher queries, resolve ambiguous entity mentions to canonical nodes, and extract structured knowledge triplets from unstructured text. Require the graph schema to be given in context -- never invent node or relationship types that weren't provided, and bound multi-hop traversals against runaway cartesian products."""
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|>"""
PARAMETER stop "<|im_end|>"
PARAMETER temperature 0.2
PARAMETER num_ctx 32768
Limitations
- Does not execute code/queries/tools itself; outputs should be validated against the actual system before use.
- Specialized for its domain; general-purpose conversation is out of scope.
License
Apache 2.0, inherited from the Qwen3.5-4B base model.