license	apache-2.0
base_model	Qwen/Qwen2.5-Coder-1.5B-Instruct
base_model_relation	quantized
datasets	mios-micro-sft
library_name	gguf
language	en
pipeline_tag	text-generation
tags	gguf llama.cpp modelpack

MiOS-Micro (1.5B)

Dedicated resident miniature model and OCI artifact for MiOS — the immutable bootc/OCI Fedora agentic operating system. Powered by Qwen2.5-Coder-1.5B-Instruct (Apache-2.0), providing sub-250ms classification, journal event triage, system command synthesis, and native OpenAI-compatible tool calling.

---

1. Overview

In an agentic operating system, background daemons and event monitors require continuous neural classification. Calling a 12B+ model for every journal error or condition check creates latency bottlenecks and burns VRAM.

MiOS-Micro solves this by maintaining a permanent, always-warm resident slot (~1.4 GB RAM / VRAM) that serves:

• mios-log-watcher.service: Real-time triage of kernel and systemd journal events.
• mios-cron-director.service: Fast condition gating before executing scheduled operations.
• prefilter / agent-pipe: Rapid capability and verb dispatch across the 100+ MiOS verbs.
• mios-micro-llm CLI: Instant natural language command-line operations.

Unlike legacy micro-models that rely on non-standard Pythonic special tokens, MiOS-Micro is fine-tuned for strict OpenAI JSON tool calling and structured outputs, adhering fully to Architectural Laws 2 and 5.

---

2. Model Architecture & Specifications

Property	Value	Rationale
Base Model	Qwen/Qwen2.5-Coder-1.5B-Instruct	State-of-the-art reasoning/coding density at 1.5B scale
License	Apache-2.0	Strict FOSS compliance (no proprietary research-only clauses)
Parameters	1.54 Billion	Capable of complex JSON schemas within edge/CPU budgets
Quantization	Q4_K_M (~0.99 GB) / Q8_0 (~1.65 GB)	Resident RAM footprint < 1.5 GB
Context Window	8,192 tokens (resident) / 32,768 native	Sized for multi-line journal streams and JSON schemas
Inference Latency	< 250 ms (GPU / AVX-512 CPU)	Immediate response for real-time daemon loops
Serving Endpoint	OpenAI /v1/chat/completions	Standard wire contract (llama-server behind llama-swap)

---

3. Four-Pillar Training Protocol

The training corpus is generated through self-distillation against the live system catalog (mios-finetune-dataset --role micro):

                       ┌───────────────────────────────────────┐
                       │       Live MiOS System Surface        │
                       │  - [verbs] catalog (100+ verbs)       │
                       │  - journald / /var/log/messages       │
                       │  - bootc / greenboot / podman specs   │
                       └──────────────────┬────────────────────┘
                                          │
                                          ▼
   ┌─────────────────────────────────────────────────────────────────────────────┐
   │                       Four-Pillar SFT Dataset                               │
   ├──────────────────────────────┬──────────────────────────────────────────────┤
   │  Pillar 1: CLI Synthesis     │  Natural language intent -> exact command    │
   │  (35% of corpus)             │  (e.g., "rollback deployment" -> "bootc      │
   │                              │   rollback")                                 │
   ├──────────────────────────────┼──────────────────────────────────────────────┤
   │  Pillar 2: Log Triage        │  Raw journal lines -> Structured JSON        │
   │  (25% of corpus)             │  (severity, subsystem, root_cause, action)   │
   ├──────────────────────────────┼──────────────────────────────────────────────┤
   │  Pillar 3: Intent Routing    │  Low-latency verb routing without full       │
   │  (25% of corpus)             │  pipeline overhead                           │
   ├──────────────────────────────┼──────────────────────────────────────────────┤
   │  Pillar 4: Function Calling  │  Strict OpenAI JSON schema validation and    │
   │  (15% of corpus)             │  two-sided tool calling conformance          │
   └──────────────────────────────┴──────────────────────────────────────────────┘

---

4. OCI Artifact Distribution

MiOS-Micro is distributed as a multi-architecture OCI Artifact and Bound Container Image conforming to the CNCF Model Distribution and OCI Image specifications.

Pulling as an OCI Container

# Run standalone micro-server on port 8500
podman run -d --name mios-micro -p 8500:8500 ghcr.io/mios-dev/mios-micro:1.5b

Pulling as a CNCF ModelPack Artifact (ORAS)

The model artifact follows ModelPack model-spec v0.0.7: artifact type application/vnd.cncf.model.manifest.v1+json, config application/vnd.cncf.model.config.v1+json, and raw weight, dataset and doc layers. It is attached to the runtime image digest as an OCI referrer and signed with cosign (keyless).

oras discover --artifact-type application/vnd.cncf.model.manifest.v1+json ghcr.io/mios-dev/mios-micro:1.5b
oras pull ghcr.io/mios-dev/mios-micro:1.5b-modelpack
# Unpacks:
#  - qwen2.5-coder-1.5b-instruct-q4_k_m.gguf   (weight.v1.raw)
#  - mios-micro-sft.jsonl                      (dataset.v1.raw)
#  - README.md, dataset-card.md                (doc.v1.raw)

The GGUF is the upstream Qwen/Qwen2.5-Coder-1.5B-Instruct-GGUF file pinned by revision and sha256 in pyproject.toml ([tool.mios-micro.package]). Until a fine-tuned GGUF replaces it, the model card declares base_model_relation: quantized; it becomes finetune at that point. The dataset is described in docs/dataset-card.md.

Binding into Bootc (Architectural Law 3)

In production MiOS images, the micro-container is registered under /usr/lib/bootc/bound-images.d/mios-micro.json, ensuring the model image is pulled during initial OS baking and updated atomically with host OS upgrades.

---

5. Development & Pipeline Quickstart

1. Environment Setup

python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"   # pyproject.toml is the single dependency source

2. Generate Dataset

# Generate the 4-pillar SFT dataset from the seed lists in src/mios_micro/dataset.py
python3 -m mios_micro.dataset --out /tmp/mios-micro-sft.jsonl

3. Fine-Tune Model

# Run LoRA fine-tuning using HuggingFace TRL SFTTrainer
python3 -m mios_micro.train \
  --base-model "Qwen/Qwen2.5-Coder-1.5B-Instruct" \
  --dataset /tmp/mios-micro-sft.jsonl \
  --output-dir ./output

4. Convert & Quantize to GGUF

python3 -m mios_micro.convert \
  --model-dir ./output \
  --quant-type Q4_K_M \
  --output mios-micro-1.5b-q4_k_m.gguf

5. Validate & Benchmark

python3 -m mios_micro.eval --model mios-micro-1.5b-q4_k_m.gguf

---

6. OpenAI API Wire Usage

Structured Log Triage

curl -s http://localhost:8500/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "mios-micro:1.5b",
    "messages": [
      {
        "role": "system",
        "content": "You are MiOS-Micro Log Triager. Output structured JSON."
      },
      {
        "role": "user",
        "content": "bootc[412]: composefs digest mismatch on commit 8f192bc"
      }
    ],
    "response_format": {"type": "json_object"}
  }' | jq .

Output:

{
  "severity": "WARN",
  "subsystem": "bootc",
  "root_cause": "composefs digest verification mismatch",
  "actionable": true,
  "recommended_action": "bootc rollback"
}

---

7. License

Distributed under the Apache-2.0 License. See LICENSE for details.