---
license: apache-2.0
base_model: Qwen/Qwen3.5-0.8B
tags:
  - qwen3.5
  - fine-tuned
  - qlora
  - json-generation
  - structured-output
  - intent-detection
  - system-commands
  - os-monitoring
language:
  - en
pipeline_tag: text-generation
library_name: transformers
---

# Qwen3.5-0.8B — OS System Command Intent Parser

Fine-tuned from [Qwen/Qwen3.5-0.8B](https://huggingface.co/Qwen/Qwen3.5-0.8B) using QLoRA on 1,059 system monitoring examples. Given a natural language system alert, the model outputs a structured JSON action plan describing intent, steps, risk level, and confidence.

---

## Example

**Input:** `RAM at 95%, system becoming unresponsive`

**Output:**

```json
{
  "intent": "memory_check",
  "steps": [
    { "action": "check_memory", "params": {} },
    { "action": "list_processes", "params": { "sort": "memory", "limit": 10 } }
  ],
  "risk": "high",
  "confidence": 0.92
}
```

---

## Supported Intents

| Intent           | Description                                |
| ---------------- | ------------------------------------------ |
| `memory_check`   | RAM pressure, OOM, swap issues             |
| `cpu_check`      | High CPU, thermal throttling, load average |
| `cleanup_disk`   | Disk full, large files, temp cleanup       |
| `disk_check`     | SMART errors, RAID, filesystem corruption  |
| `network_check`  | Interface down, DNS, latency, VPN          |
| `service_check`  | Service crashed, restart needed            |
| `process_check`  | Zombie, runaway, file descriptor limits    |
| `security_alert` | Brute force, intrusion, exploit detected   |
| `system_check`   | Kernel panic, hardware errors, reboot      |
| `diagnose`       | Multi-condition or ambiguous alerts        |
| `reject`         | Non-system inputs, dangerous commands      |

---

## Output Schema

| Field        | Type   | Description                 |
| ------------ | ------ | --------------------------- |
| `intent`     | string | Alert category              |
| `steps`      | array  | Ordered actions with params |
| `risk`       | string | `low` / `medium` / `high`   |
| `confidence` | float  | Model confidence 0–1        |

---

## Training Details

| Property            | Value                       |
| ------------------- | --------------------------- |
| Base model          | Qwen/Qwen3.5-0.8B           |
| Method              | QLoRA (LoRA r=16, α=16)     |
| Dataset             | 1,059 custom JSONL examples |
| Epochs              | 3                           |
| Optimizer           | adamw_8bit                  |
| Learning rate       | 2e-4                        |
| LR scheduler        | cosine                      |
| Max sequence length | 1024                        |
| Hardware            | Tesla T4 — Google Colab     |

---

## Prompt Format

Uses Qwen3.5's native ChatML format:

```
<|im_start|>system
You are a structured data extractor. Given an input, output a valid JSON object and nothing else.<|im_end|>
<|im_start|>user
{system alert}<|im_end|>
<|im_start|>assistant
```

---

## GGUF — Local Inference

```bash
# Download and run (auto-downloads on first run)
llama-server -hf exoro/qwen3.5:Q4_K_M --reasoning-budget 0 --port 8080

# Test
curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3.5",
    "messages": [
      {"role": "system", "content": "You are a structured data extractor. Given an input, output a valid JSON object and nothing else."},
      {"role": "user", "content": "RAM at 95%, system becoming unresponsive"}
    ]
  }'
```

---

## Limitations

- Trained on system-monitoring domain — may not generalise to unrelated tasks
- Risk and confidence values are heuristic, not calibrated
- Small model (0.8B) — complex multi-step plans may occasionally be incomplete

---