---
license: apache-2.0
base_model: Qwen/Qwen2.5-3B-Instruct
tags:
- opnsense
- firewall
- network-security
- function-calling
- onnx
- onnxruntime-genai
- cpu-inference
- qwen2.5
pipeline_tag: text-generation
library_name: onnxruntime-genai
---

# OPNsense Agent — Qwen2.5-3B ONNX int4 (CPU)

Modèle **Qwen2.5-3B-Instruct** fine-tuné (LoRA) sur le domaine **Firewall OPNsense**,
mergé et quantifié en **ONNX int4** pour inférence CPU sans GPU.

Fait partie du projet [cyber-agent-engine](https://github.com/patlegu/cyber-agent-engine) —
architecture multi-agents cybersécurité avec protocole CAP v1.

## Performances

| Métrique | Valeur |
|---|---|
| Score validation | **102/102 (100%)** |
| Fonctions couvertes | 102 |
| Format | ONNX int4 (onnxruntime-genai) |
| Base model | Qwen/Qwen2.5-3B-Instruct |
| Latence CPU (8 vCPU) | ~8–12s/requête |

## Usage

```python
import onnxruntime_genai as og
import json

model = og.Model("patlegu/opnsense-qwen25-onnx-int4")
tokenizer = og.Tokenizer(model)
tokenizer_stream = tokenizer.create_stream()

cap = {
  "directive": "block_ip",
  "entities": {"IP_ADDRESS": ["203.0.113.42"], "INTERFACE": ["wan"]},
  "context": {"source": "coordinator", "confidence": 0.97}
}

prompt = (
    "<|im_start|>system\n"
    "Tu es un agent OPNSENSE. Tu reçois des directives structurées du coordinateur "
    "sous forme de paquets JSON (format CAP v1) et tu génères des appels d'API précis "
    "sous forme de tool_calls. Tu ne réponds jamais en langage naturel.\n"
    "<|im_end|>\n"
    "<|im_start|>user\n"
    + json.dumps(cap) +
    "\n<|im_end|>\n"
    "<|im_start|>assistant\n"
)

import numpy as np
input_ids = np.array(tokenizer.encode(prompt), dtype=np.int32)

params = og.GeneratorParams(model)
params.set_search_options(max_length=len(input_ids) + 256, temperature=0.1, do_sample=False)

generator = og.Generator(model, params)
generator.append_tokens(input_ids)

output_tokens = []
while not generator.is_done():
    generator.generate_next_token()
    token = generator.get_next_tokens()[0]
    output_tokens.append(token)

print(tokenizer.decode(output_tokens))
# → [{"type": "function", "function": {"name": "...", "arguments": "..."}}]
```

## Installation

```bash
pip install onnxruntime-genai numpy
```

## Format d'entrée (CAP v1)

Agent OPNsense capable de traduire des directives structurées (protocole CAP v1) en appels API OPNsense précis. Couvre 102 fonctions : règles firewall, aliases, NAT, IDS Suricata, Traffic Shaping, ACME/TLS, IPsec, OpenVPN, diagnostics.

L'agent reçoit un paquet JSON structuré :

```json
{
  "directive": "block_ip",
  "entities": {"IP_ADDRESS": ["203.0.113.42"], "INTERFACE": ["wan"]},
  "context": {"source": "coordinator", "confidence": 0.97}
}
```

Et produit un `tool_call` JSON :

```json
[{"type": "function", "function": {"name": "directive_name", "arguments": "{}"}}]
```

## Projet

- Repo principal : [cyber-agent-engine](https://github.com/patlegu/cyber-agent-engine)
- Architecture : coordinateur Qwen2.5-3B + agents SLM LoRA + protocole CAP v1
- Autres agents : [OPNsense](patlegu/opnsense-qwen25-onnx-int4) · [WireGuard](patlegu/wireguard-qwen25-onnx-int4) · [CrowdSec](patlegu/crowdsec-qwen25-onnx-int4)