OCC Pillar 2 · The Open AI Governance Language Model
compliance-qwen
Qwen2.5-7B · QLoRA · Apache-2.0 · EU-hosted
The first open, reproducible, EU-hosted language model for AI Governance. QLoRA-finetuned on the OCC Corpus — no proprietary data, no cloud training, Apache-2.0 licensed.
Technical Specification
base Qwen/Qwen2.5-7B-Instruct
method QLoRA (rank=64, alpha=128)
data 689 Chunks from 26 verified sources
training pairs 241
license Apache-2.0 (Base) + CC-BY 4.0 (Data)
compute IIO AI Hub — RTX PRO 6000 (96GB VRAM)
target Q4 2026
Development Progress
5/11 Milestones · 45%
OCC Corpus (689 Chunks from 26 Sources)
Eval-Sets (50 Questions, 7 Categories)
Red-Team-Playbook (7 Scenarios, 8 Gates)
chunks-to-training.py (241 Alpaca Pairs)
train_lora.py (QLoRA, dry-run OK)
Pilot Training Run (1 Epoch)
Eval: Grade B+ (pass_rate ≥ 0.65)
Red-Team: all 8 Gates green
Model Card + System Card
Hugging Face Upload (HITL)
LiteLLM Integration (as "compliance-qwen")
Training Data
compliance-qwen is trained exclusively on verified public-domain sources and the IIO Framework (Apache-2.0). No proprietary data, no copyright violations, no personal data.
241 Alpaca-Format Pairs generated from 689 chunks
2 Strategies Definition + Explanation
DE/EN Language German primary
Public Domain + Apache-2.0 Licenses 100% verified
Safety & Evaluation
No release without passing all 8 safety gates. The OCC Eval & Safety Lab tests before every release:
RT-001 Jailbreak Protection 0 successes in 20 attempts
RT-002 PII Protection No extractable personal data
RT-003 MCP Scope 100% tool scope enforced
RT-004 Rate Limit Burst test passed
RT-005 CBRN Protection 0 harmful responses
RT-006 Hallucination <10% on eval set
RT-007 Attribution 100% source attribution
RT-008 Eval Grade ≥ Grade B (pass_rate ≥ 0.65)
Currently Available: OCC AI Assistant
While compliance-qwen is in development, the OCC AI Assistant is available — grounded on 689 chunks, EU-hosted, anonymously usable.