RingoSystems/ringollm-v42

Windows endpoint management specialist for DSC v3, PowerShell 7 and Intune. A Qwen3.6-27B fine-tune that emits schema-valid DSC v3 and refuses out-of-scope questions instead of inventing answers.

A LoRA fine-tune of Qwen3.6-27B specialised for Windows endpoint management, quantized to Q5_K_M (19 GB).

In scope: DSC v3, DSC Legacy, PowerShell 7, Intune deployment (Settings Catalog, OMA-URI, IME, Win32 packaging, Proactive Remediations), and endpoint-side Entra ID diagnostics.

Deliberately out of scope: Graph SDK administration, KQL/Sentinel, WDAC/App Control, Cisco and network appliances, Linux, macOS. The model is trained to refuse these rather than guess, and it will.

Quick start

ollama run ringosystems/ringollm-v42:4.2

The model ships with its system prompt baked in. Send no system message unless you intend to replace the scope definition and refusal behaviour entirely.

Why this version exists

Earlier builds emitted securityContext: Elevated in DSC v3 documents. That is schema-invalid. dsc-lib declares the enum as:

#[serde(rename_all = "camelCase")]
pub enum SecurityContextKind { Current, Elevated, Restricted }

On single-word variants rename_all = "camelCase" accepts only the lowercase form. The JSON schema agrees (enum: ["current","elevated","restricted"]), as do DSC’s own shipped examples. Microsoft Learn’s ValidValues table lists the capitalised spellings and is wrong — that table is the origin of the bug.

Measured on a 25-question rubric (129 checks, greedy decoding), against the previous release:

	previous	v4.2
rubric checks	¹¹⁰⁄₁₂₉ (85.3%)	¹¹²⁄₁₂₉ (86.8%)
questions fully passed	¹⁴⁄₂₅	¹⁷⁄₂₅
`securityContext` valid / invalid	0 / 6	8 / 0
out-of-scope refusals	⁴⁄₅	⁵⁄₅

The previous release emitted an invalid securityContext on every DSC document it produced — six for six, never once correct. This one is eight for eight correct.

Example

Prompt: Write a minimal DSC v3 YAML configuration that requires elevation.

$schema: https://aka.ms/dsc/schemas/v3/bundled/config/document.json
metadata:
  Microsoft.DSC:
    securityContext: elevated
resources: []

Note the lowercase elevated.

Defaults

Baked into the model, override as needed:

parameter	value
`temperature`	0.3
`top_p`	0.9
`repeat_penalty`	1.15
`repeat_last_n`	256
`num_ctx`	65536
`num_predict`	2048

Training data

Q&A pairs are derived from Microsoft documentation, the DSC resource registry and the published JSON schemas via deterministic builders. No LLM-generated text is present in the training data — reasoning traces were removed before training, which also matches how the model is served (reasoning suppressed).

Limitations

Scoped to Windows endpoint management; it refuses elsewhere by design.
Reasoning is suppressed — the chat template pre-closes the <think> block.
Q5_K_M quantization; expect small deviations from the full-precision merge.
Knowledge is bounded by the training corpus. Verify anything load-bearing against current Microsoft documentation, and verify enum values against the DSC source or JSON schema rather than Microsoft Learn.

License

Apache 2.0, inherited from Qwen3.6-27B. Built from the unsloth mirror.