AOS-COS-v1.0 — ArmadaOS Chief of Staff

Custom fine-tuned model for the ArmadaOS Chief of Staff agent. Built on Qwen3.5-9B with 262K native context window.

Model Details

• Base Model: Qwen3.5-9B (262K context, extensible to 1M with YaRN)
• Fine-tuning Method: bf16 LoRA (rank 16, alpha 16)
• Training Data: 315 Gold Standard examples across 7 capability categories
  • SFT: 220 examples (155 trajectory + 65 function calling)
  • DPO: 95 preference pairs
• Quantization: Q4_K_M (5.02 bits per weight)
• File Size: 5.3 GB

Capability Categories

Category	Description	Examples
A	Boot & Identity	50
B	Error Correction Protocol	15
C	10-Layer Memory System	50
D	Operational Cadence	50
E	Governance & Constitution	50
F	Failure & Antifragility	50
G	Decision & Communication	50

Usage with Ollama

ollama run ArmadaOS/AOS-COS-v1.0

Training Infrastructure

• GPU: NVIDIA RTX A6000 (48GB) / L40 (48GB)
• Platform: RunPod
• Framework: Unsloth + TRL
• Training Time: ~45 min SFT + ~20 min DPO

Version History

• v1.0 — Initial release. SFT + DPO on 315 Gold Standard v5.3 examples.

---

Built by ArmadaOS. Compound, Never Lose.