92 2 months ago

Your Windows endpoint management badass. A fine-tuned Qwen 3.5 (26.9B) built for Microsoft Intune deployment, Desired State Configuration, and PowerShell 7. It answers with real cmdlets, scripts, and configuration, and it refuses cleanly when unknown.

ollama run RingoSystems/ringollm

Models

View all →

Readme

RingoLLM

Your endpoint management badass.

Fine-tuned Qwen 3 8B for Windows endpoint management. Provides precise, actionable answers with real cmdlets, scripts, registry paths, and configuration examples.

Quick Start

ollama run RingoSystems/ringollm

Specializations

  • Microsoft Intune — device configuration, compliance policies, Win32 app deployment, Proactive Remediations, Settings Catalog, IME troubleshooting
  • PowerShell 7 — Graph SDK, module development, secrets management, parallel processing, enterprise automation
  • DSC v3 — YAML configuration authoring, resource schemas, Registry/Service/WinGet resources, Azure Machine Configuration
  • Microsoft Graph API — device management endpoints, batch requests, certificate authentication, pagination
  • Entra ID — Conditional Access, device identity lifecycle, stale device cleanup
  • Security Baselines — CIS benchmarks, Defender AV, BitLocker, firewall hardening, security policy auditing
  • WinGet — package management, private REST sources, DSC integration
  • Azure Monitor — Data Collection Rules, Log Analytics, Windows Event Log collection

Model Details

Property Value
Base Model Qwen 3 8B
Version v3.02
Fine-tuning QLoRA (rank 64, alpha 64) via Unsloth
Training Data 8,135 curated pairs + 110 gold pairs (2x weighted)
Epochs 2
Quantization Q5_K_M (~5.6 GB)
Context Length 8192 tokens
Thinking Mode Disabled
License Apache 2.0

Evaluation

Metric Score
Eval pass rate 1230 (40%)
Eval raw score 121150 (81%)
Benchmark clean stops 5050 (100%)
Avg response time 2.4s

Version History

Version Base Model Pass Rate Raw Score
v1.0 Qwen 2.5 Coder 7B 730 (23%) 74150 (49%)
v3.0 Qwen 3 8B 430 (13%) 90150 (60%)
v3.02 Qwen 3 8B 1230 (40%) 121150 (81%)

Usage

All inference parameters (stop tokens, chat template, system prompt) are pre-configured. No manual setup required.

ollama run RingoSystems/ringollm

Important Notes

The model uses Qwen 3 with thinking mode disabled. The chat template includes a <think> block that forces direct answers without chain-of-thought reasoning. Do NOT add <|im_start|> as a stop token — Qwen 3 generates this internally.

Example Prompts

  • “Write a PowerShell script to get all non-compliant Intune devices using Microsoft Graph”
  • “Create a DSC v3 YAML configuration that sets AllowTelemetry to 0 via the registry”
  • “How do I set up Conditional Access to require compliant devices for M365 apps?”
  • “Write a PowerShell function with retry logic for Graph API throttling (429)”
  • “Explain the difference between DSC v3 resource types: built-in, PowerShell adapter, and command-based”

Training Data

Curated from 92,000+ synthetic Q&A pairs generated from 26,273 public documentation sources:

  1. Merge and deduplicate — 79,940 pairs from 50 source files
  2. Domain removal — off-topic content stripped
  3. Content repair — hallucinated CLI commands and properties fixed
  4. Cmdlet validation — 318 hallucinated cmdlets removed via ground-truth lookup (2,465 Graph SDK cmdlets, 37 PS cmdlets, 17 DSC v3 resources)
  5. Domain capping — balanced to 8,135 pairs across 11 domains
  6. Gold pairs — 110 hand-validated pairs weighted 2x at training time

All training data sourced exclusively from publicly available Microsoft documentation and community repositories.

Known Limitations

  • 8B parameter model; complex multi-step reasoning may be less accurate than larger models
  • Graph SDK module names and specific cmdlet names are the weakest area (70% raw score)
  • Training data has a knowledge cutoff; very recent API changes may not be reflected
  • PowerShell code should be reviewed before production deployment
  • Focused on Windows 1011 client endpoint management; limited server/infrastructure coverage

Links