A specialized fine-tune of Qwen3.6-27B focused on Windows endpoint management. RingoLLM produces precise, actionable answers about DSC v3, DSC Legacy, PowerShell 7, Microsoft Intune, Entra ID, and security configuration baselines — and refuses cleanly on out-of-scope topics rather than fabricating syntax.

What it's good at

Domain	Eval pass	Eval score
PowerShell 7	6/6 (100%)	21/21 (100%)
DSC v3	7/10 (70%)	69/73 (95%)
Out-of-scope refusals	3/5 (60%)	27/29 (93%)
Intune endpoint deployment	1/2 (50%)	7/8 (88%)
Overall	17/25 (68%)	139/154 (90%)

Eval harness: 25 questions covering DSC v3, DSC Legacy, Intune, PowerShell 7, and out-of-scope refusal. Avg response time 13s.

What it produces

DSC v3 YAML in canonical form:

camelCase property names (keyPath, valueName, valueData)
PascalCase enum values (securityContext: Elevated)
Typed value wrappers (DWord, String, Binary, etc.)
_exist: true|false for state, _ensure is rejected
Windows-style registry paths (HKLM\Path, no HKLM:)

PowerShell 7 with proper cmdlet preference (Get-CimInstance over Get-WmiObject), correct profile paths, idempotent patterns.

Intune Settings Catalog and OMA-URI configurations with valid policy keys.

What it refuses

Cisco/network appliance configuration, Linux administration, KQL/Sentinel queries, WDAC/App Control, Microsoft Graph SDK administration, macOS — these all return brief refusals with vendor doc pointers, not fabricated syntax.

Files in this repo

File	Purpose	Size
ringollm-v411-Q5_K_M.gguf	Production-quality GGUF for Ollama / llama.cpp	18 GB
ringollm-v411-Q4_K_M.gguf	Smaller GGUF for VRAM-constrained inference	16 GB
adapter_model.safetensors	LoRA adapter (for further fine-tuning from base)	305 MB
Modelfile.ringollm-v4	Ollama Modelfile with proper ChatML template	<1 KB

Quick start (Ollama)

bash

ollama pull ringosystems/ringollm
ollama run ringosystems/ringollm "Write a DSC v3 config to disable LLMNR"

Quick start (llama.cpp)

bash

./llama-server -m ringollm-v411-Q5_K_M.gguf \
  -c 8192 \
  --chat-template-file chat_template.jinja

Ollama Modelfile

This model uses a Qwen3.6 ChatML template with thinking mode disabled. The Modelfile.ringollm-v4 in this repo includes the canonical configuration. Do not use TEMPLATE {{ .Prompt }} — it strips the system prompt and disables refusal behavior.

Training details

Base: Qwen3.6-27B
Method: LoRA (r=16, alpha=16, dropout=0)
Target modules: q/k/v/o/gate/up/down projections (no modules_to_save)
Trainable params: 79.7M / 27.4B (0.29%)
Corpus: 2,539 effective training pairs (filtered v4 corpus + firm-refusal pairs + adversarial pairs + in-scope corrections, replicated per category)
Hyperparameters: lr=1e-4, batch=1x4 grad accum, 1.5 epochs, warmup 10 steps, early stopping patience 2
Hardware: AMD Ryzen AI Max (Strix Halo) with Radeon 8060S iGPU, ROCm 7.13, Unsloth 2026.4.8
Wall-clock: 3.6 hours, early-stopped at step 700 (best eval_loss 0.7538)

Known limitations

Legacy DSC (1.1) generation is unreliable. The model has internalized DSC v3 syntax so deeply that legacy DSC questions often produce v3 output instead. If you need DSC 1.1 specifically, this isn't the right model.
Eval criteria are conservative. Several "fails" in the harness are regex specificity issues rather than model failures. Real-world correctness rate is closer to 80%.
No tool calling. The base model supports it; the fine-tune does not preserve tool-calling capability cleanly.

License

Apache 2.0 (inherits from Qwen3.6-27B base).

Citation

If you use RingoLLM in your work or build on this fine-tune, attribution to RingoSystems is appreciated but not required.

RingoSystems/RingoLLM

RingoLLM v4