---

language:

• en
• de

license: apache-2.0

base_model: Qwen/Qwen3.8-27B

tags:

• compound-ai
• judge
• quality-assurance
• evaluation
• paraconsistent
• reasoning
• gguf
• lumi-g
• moe-sovereign

datasets:

• moe-sovereign/judge-evaluation-sft

pipeline_tag: text-generation

library_name: transformers

---

⚖️ MoE Sovereign Judge 27B GGUF (moe-sovereign-judge-27b)

High-Precision Compound-AI Quality Evaluation, Paraconsistent Reconciliation & Zero-Fallback Judge




---

📌 Executive Summary & Architectural Role

moe-sovereign-judge-27b is a specialized 27-billion parameter LLM fine-tuned on the EuroHPC LUMI-G Supercomputer (8× AMD Instinct™ MI250X GCDs (4× physical modules, 64GB HBM2e per GCD)) for high-assurance, paraconsistent quality evaluation within the MoE Sovereign compound-AI platform.

Built on top of Qwen3.8-27B, this model serves as the primary Sovereign Judge responsible for:

1. Semantic Code & Solution Verification: Evaluating complex systems code (C++20 lock-free data structures, Linux eBPF/XDP filters, Rust memory safety).
2. Paraconsistent Knowledge Reconciliation: Reconciling conflicting retrieved knowledge nodes, provenance data, and temporal policy updates without hallucinating.
3. Structured JSON Output: Producing strict, schema-compliant JSON evaluation verdicts containing quality_score, factuality_score, overall_score, and detailed analytical rationale.

This model directly replaces the previous sovereign-judge:35b-q4km checkpoint, eliminating socket timeouts, lowering VRAM requirements, and completely eliminating UNSCORED_FALLBACK verdicts.

---

⚙️ Model Details & Training Infrastructure

• Base Model: Qwen/Qwen3.8-27B
• Training Supercomputer: EuroHPC LUMI-G (CSC Finland)
• Hardware Allocation: 1 Node (8× AMD Instinct™ MI250X GCDs, 4× physical modules, 64GB HBM2e per GCD, 112 CPU cores)
• Job Execution: Slurm Job ID 21263413 (moe_expert_pipe)
• Training Duration: 24 hours, 39 minutes, 57 seconds (1-00:39:57)
• Training Methodology: Supervised Fine-Tuning (SFT) + Direct Preference Optimization (DPO) on verified MoE Sovereign execution traces.
• Format & Quantization: GGUF (Q4_K_M), 262144 context length.

---

📊 Benchmark Verification Highlights

Evaluated against the MoE Sovereign Scientific Multidisciplinary Benchmark (August 2026):

| Benchmark Category | Task | Judge Verdict Score | Deterministic Verification |

| :--- | :--- | :---: | :---: |

| Paraconsistent Knowledge | sci-graphrag-02 (Graph Reconciliation) | 9.4 / 10.0 🌟 | 10.0 / 10.0 (100% Factually Grounded) |

| Systems Programming | sci-sysprog-01 (Lock-Free MPSC Ring Buffer) | 7.0 / 10.0 | 10.0 / 10.0 (100% Memory Barrier Compliant) |

| Network Infrastructure | sci-sysprog-02 (eBPF XDP Map Sync) | 7.0 / 10.0 | 10.0 / 10.0 (100% Kernel Verified) |

| Governance & Sovereignty | sci-governance-01 (Technical Sovereignty) | 6.2 / 10.0 | 8.0 / 10.0 (Policy Grounded) |

---

🚀 Deployment & Ollama Integration

1. Modelfile Configuration

Create a file named Modelfile:

FROM ./moe-sovereign-judge-27b-q4km.gguf

TEMPLATE """{{ if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}{{ if .Prompt }}<|im_start|>user
{{ .Prompt }}<|im_end|>
{{ end }}<|im_start|>assistant
"""

PARAMETER stop "<|im_start|>"
PARAMETER stop "<|im_end|>"
PARAMETER temperature 0.1
PARAMETER num_ctx 262144
PARAMETER num_predict 512

SYSTEM """You are the MoE Sovereign Scientific Quality Judge. Evaluate the provided solution strictly against factual, technical, and mathematical correctness requirements. Respond ONLY with a valid JSON object."""

2. Import into Ollama

# Pull model directly or build from GGUF
ollama create sovereign-judge:27b-q4km -f Modelfile

---

📜 Citation & License

• License: Apache 2.0
• Repository: https://github.com/h3rb3rn/moe-sovereign
• Documentation: https://docs.moe-sovereign.org