Instructions to use h3rb3rn/moe-sovereign-judge-27b-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use h3rb3rn/moe-sovereign-judge-27b-GGUF with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="h3rb3rn/moe-sovereign-judge-27b-GGUF")# Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("h3rb3rn/moe-sovereign-judge-27b-GGUF", device_map="auto") - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use h3rb3rn/moe-sovereign-judge-27b-GGUF with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "h3rb3rn/moe-sovereign-judge-27b-GGUF" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "h3rb3rn/moe-sovereign-judge-27b-GGUF", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'Use Docker
docker model run hf.co/h3rb3rn/moe-sovereign-judge-27b-GGUF
- SGLang
How to use h3rb3rn/moe-sovereign-judge-27b-GGUF with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "h3rb3rn/moe-sovereign-judge-27b-GGUF" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "h3rb3rn/moe-sovereign-judge-27b-GGUF", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "h3rb3rn/moe-sovereign-judge-27b-GGUF" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "h3rb3rn/moe-sovereign-judge-27b-GGUF", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }' - Docker Model Runner
How to use h3rb3rn/moe-sovereign-judge-27b-GGUF with Docker Model Runner:
docker model run hf.co/h3rb3rn/moe-sovereign-judge-27b-GGUF
βοΈ MoE Sovereign Judge 27B GGUF (moe-sovereign-judge-27b)
High-Precision Compound-AI Quality Evaluation, Paraconsistent Reconciliation & Zero-Fallback Judge
π Executive Summary & Architectural Role
moe-sovereign-judge-27b is a specialized 27-billion parameter LLM fine-tuned on the EuroHPC LUMI-G Supercomputer (8Γ AMD Instinctβ’ MI250X GCDs (4Γ 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:
- Semantic Code & Solution Verification: Evaluating complex systems code (C++20 lock-free data structures, Linux eBPF/XDP filters, Rust memory safety).
- Paraconsistent Knowledge Reconciliation: Reconciling conflicting retrieved knowledge nodes, provenance data, and temporal policy updates without hallucinating.
- 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.
βοΈ Model Details & Training Infrastructure
- Base Model:
Qwen/Qwen3.8-27B - Training Supercomputer: EuroHPC LUMI-G (CSC Finland)
- Hardware Allocation: 1 Node (8Γ AMD Instinctβ’ MI250X GCDs, 4Γ 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.
π 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 π |
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) |
π 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
π Citation & License
- License: Apache 2.0
- Repository: https://github.com/h3rb3rn/moe-sovereign
- Documentation: https://docs.moe-sovereign.org
Model tree for h3rb3rn/moe-sovereign-judge-27b-GGUF
Base model
Qwen/Qwen3.8-27B