ringollm-v42

A LoRA fine-tune of unsloth/Qwen3.6-27B specialised for Windows endpoint management: DSC v3, DSC Legacy, PowerShell 7, Intune deployment (Settings Catalog, OMA-URI, IME, Win32 packaging, Proactive Remediations) and endpoint-side Entra ID diagnostics.

Quantized to Q5_K_M for llama.cpp.

Why this exists

Earlier releases emitted securityContext: Elevated in DSC v3 documents. That is schema-invalid. dsc-lib's SecurityContextKind is declared #[serde(rename_all = "camelCase")], so only current, elevated and restricted deserialize — the PascalCase form fails outright. Microsoft Learn's ValidValues table lists the capitalised spellings and is wrong. This release emits the lowercase enum.

Measured results

metric	ringollm-v4	ringollm-v42
rubric checks passed	110/129 (85.3%)	112/129 (86.8%)
questions passed	14/25 (56.0%)	17/25 (68.0%)
securityContext valid / invalid	0 / 6	8 / 0
median tok/s	11.26	11.29

Usage

llama.cpp:

llama-server -m RingoLLM-v42-qwen35-27B-Q5_K_M.gguf \
  --jinja --chat-template-file ringollm-v42.jinja \
  -ngl 999 -fa on -np 1 \
  -c 65536 --cache-type-k q8_0 --cache-type-v q8_0 --kv-unified \
  --temp 0.3 --top-p 0.9 --repeat-penalty 1.15 --repeat-last-n 256

Ollama:

ollama create ringollm-v42 -f Modelfile
ollama run ringollm-v42

Serving parameters, and why

flag	value	why it matters
--chat-template-file	ringollm-v42.jinja	Required. Ships in this repo. It injects the default system prompt that defines scope and the refusal format. Without it the model loses all refusal behaviour.
--kv-unified	on	Stability, not tuning. On RADV / AMD Strix Halo this architecture hard-hangs the GPU on prompts past ~20k tokens without it — vk::Queue::submit: ErrorDeviceLost, an amdgpu ring timeout, and a reset that takes down unrelated processes. Upstream issue closed as not-planned, so there is no fix to wait for. Harmless on other backends; leave it on.
--cache-type-k/v	q8_0	Pairs with --kv-unified to make long contexts affordable and stable.
-c	65536	Observed real-world high-water is ~26k tokens.
-np	1	Recurrent state is allocated per slot; more slots divide the usable context.
-fa	on	Flash attention.
--temp / --top-p	0.3 / 0.9	What the model was evaluated with.
--repeat-penalty / --repeat-last-n	1.15 / 256	As evaluated.

Prompting

The chat template injects a default system prompt when the caller supplies none. It defines the scope and the refusal format for out-of-scope questions, so omit the system message unless you intend to replace that behaviour entirely.

Training data

All training 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.

Limitations

• Scoped to Windows endpoint management. It is instructed to refuse elsewhere, and will.
• Served with reasoning suppressed; the template pre-closes the <think> block.
• Quantized to Q5_K_M; expect small deviations from the full-precision merge.