---
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"}
    ]
  }'
```

| File | Size | Use case |
|---|---|---|
| `Qwen3.5-0.8B.Q4_K_M.gguf` | ~505 MB | Best balance |
| `Qwen3.5-0.8B.Q8_0.gguf` | ~774 MB | Higher quality |
| `Qwen3.5-0.8B.F16.gguf` | ~1.6 GB | Full precision |

---

## 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

---

## Related Repos

| Repo | Contents |
|---|---|
| [exoro/qwen3.5](https://huggingface.co/exoro/qwen3.5) | This repo — merged model + GGUF |
| [exoro/qwen3.5-0.8B-finetuned-merged](https://huggingface.co/exoro/qwen3.5-0.8B-finetuned-merged) | Earlier checkpoint |