RingoSystems/ ringollm-v51:5.1

2 3 weeks ago

A Muse Glimmer 30B fine-tune for Windows endpoint management: DSC v3, PowerShell 7, and Intune. Matches a dedicated domain fine-tune on accuracy while keeping full tool calling. Refuses legacy DSC 1.1 by design. 

ollama run RingoSystems/ringollm-v51:5.1

Details

3 weeks ago

978f74b87fdd · 17GB ·

model
muse-glimmer
·
27.9B
·
Q4_K_M
system
You are RingoLLM v5.1, a specialized AI assistant for Windows endpoint management. You focus on DSC
license
Apache License 2.0 RingoLLM v5.1 is a LoRA fine-tune of Meta Muse Glimmer 30B (meta-models/Muse-Glim
params
{ "num_ctx": 65536, "num_predict": 2048, "stop": [ "<|eot|>" ], "tempera
messages
[{"role":"user","content":"Write a minimal DSC v3 YAML configuration that requires elevation."},{"ro

Readme

RingoLLM v5.1 — Windows endpoint management (DSC v3, PowerShell 7, Intune)

A LoRA fine-tune of Meta Muse Glimmer 30B for Windows endpoint management, built to be usable as an agent: it keeps the base model’s tool calling fully intact while learning DSC v3.

Which model should you use

Your work Use
DSC v3, PowerShell 7, Intune, agentic / tool-calling workflows RingoLLM v5.1 (this model)
Legacy DSC 1.1 (Configuration blocks, MOF, Start-DscConfiguration) Not this model. Use RingoLLM v4.2 or a general-purpose model

This model refuses legacy DSC by design.

What makes this different

Narrow fine-tunes for this domain usually lose tool calling entirely. The previous model in this line scored 1 of 5 on tool selection — and that single point is an abstention case it passes only because it never calls a tool at all. Accurate on DSC, useless as an agent.

v5.1 scores 5⁄5 on tool selection and 15⁄15 on argument extraction, matching the stock base model, while raising DSC v3 accuracy from 56% to 89%.

Scored on 23 questions covering DSC v3, PowerShell, Intune and out-of-scope refusal. Legacy DSC is excluded from the comparison because this model is designed to decline it.

v5.1 prior fine-tune base model
Core rubric 90.1% 90.8% 70.2%
DSC v3 89% 90% 56%
Intune 88% 62% 75%
PowerShell 95% 95% 100%
Tool selection 5⁄5 1⁄5 5⁄5

It matches the dedicated domain fine-tune on accuracy — within one point out of 131 — and adds working tool use.

The trade

Training on a DSC v3 corpus caused catastrophic forgetting of legacy DSC 1.1 (39%, below the base model’s 74%). Rather than answer unreliably, the bundled system prompt makes the model refuse legacy DSC and offer the DSC v3 equivalent. Converting legacy DSC into v3 is in scope; producing legacy DSC is not.

Behaviour notes

  • It always reasons. The reasoning channel cannot be disabled; reasoning_strength (low/medium/high/xhigh) only changes how much. Reasoning counts against the response budget, so a small num_predict can return empty content. The bundled default is 2048.
  • Recommended sampling is preset: temperature 1.0, top_p 0.95, top_k 64. These are Meta’s published values for the base and are higher than a typical fine-tune wants — leave them.
  • 64K context configured by default; the base supports more.

Usage

ollama run ringollm-v5.1 "Write a DSC v3 YAML config that disables SMBv1."

Endpoint note

Use the chat API (ollama run, /api/chat, or the OpenAI-compatible /v1/chat/completions). These render this model correctly.

The legacy raw-completion endpoint (/api/generate) does not yet apply this model’s channel format correctly in Ollama and will prefix responses with a stray to=user marker. That is an Ollama templating limitation for this newly-released architecture, not a model defect – verified by driving the same weights with an exact raw prompt, which returns clean output.

License and attribution

Apache 2.0, inherited from Muse Glimmer 30B by Meta. Training corpus is private and not redistributed.

GGUF weights, LoRA adapter, and full benchmark methodology: https://huggingface.co/RingoSystems/RingoLLM-v5.1-MuseGlimmer-30B-GGUF