README.md

---
language:
- en
license: llama3.2
library_name: optimum
pipeline_tag: text-classification
tags:
- prompt-guard
- onnx
- onnxruntime
- mobile
- cybersecurity
- phishing-detection
- prompt-injection
- text-classification
base_model: meta-llama/Llama-Prompt-Guard-2-86M
base_model_relation: finetune
datasets:
- naserabdullahalam/phishing-email-dataset
- marycamilainfo/prompt-injection-malignant
---

# Prompt Guard 2 Multitask ONNX

This repository contains an **ONNX export** of a multitask security classifier built from **`meta-llama/Llama-Prompt-Guard-2-86M`**.

The model was fine-tuned as a **single multitask adapter** on two security-focused tasks and then merged into a standalone model before ONNX export.

## Base model

- **Base model:** `meta-llama/Llama-Prompt-Guard-2-86M`
- **Architecture:** sequence classification
- **Export format:** ONNX
- **Primary runtime:** ONNX Runtime / ONNX Runtime Mobile

## Tasks

This model is intended to score text as **BENIGN** or **MALICIOUS** across two security-related input types:

1. **Phishing email detection**
2. **Prompt injection detection**

The model uses a shared binary label space:

- `BENIGN`
- `MALICIOUS`

## Training data

This multitask model was trained using data derived from:

- `naserabdullahalam/phishing-email-dataset`
- `marycamilainfo/prompt-injection-malignant`

Additional benign prompt-style examples were included so the prompt-injection side of the multitask classifier had both positive and negative examples.

## Input format

During training and inference, inputs are prefixed with a simple modality tag:

- `[EMAIL] ...`
- `[PROMPT] ...`

### Example inputs

#### Phishing email example

```text
[EMAIL] Subject: Verify your payroll account now. Body: Your payroll access will be suspended unless you confirm your credentials here: http://example-login-reset.com