---
license: apache-2.0
tags:
- prompt-injection
- jailbreak-detection
- guardrails
- security
- llm-security
- ai-safety
- ai-agent-security
- patronus
- multilingual
- modernbert
- onnx
- quantized
- edge
base_model: patronus-studio/wolf-defender-prompt-injection
pipeline_tag: text-classification
language:
- de
- en
---

# Model Card for Wolf Defender Prompt-Injection Detector Edge

**Edge/quantized ONNX variant of [`patronus-studio/wolf-defender-prompt-injection`](https://huggingface.co/patronus-studio/wolf-defender-prompt-injection).** Private / internal.

Quantized with the Patronus RunPod ONNX method (`quantize_dynamic` INT8 linears + `MatMulNBitsQuantizer`
4-bit block-128 embedding). The parent FP32/FP16 model is unchanged; this repo only adds the small variants.
Binary prompt-injection detector (mmBERT-base).

## Benchmark — Qualifire

Evaluated on [Qualifire](https://huggingface.co/datasets/rogue-security/prompt-injections-benchmark)
(5,000 prompts, max_length 2048). Quantization is lossless vs FP32.

| Variant | Size | F1 | Precision | Recall | FPR | Accuracy |
|---|---|---|---|---|---|---|
| fp32 (parent) | — | 0.9395 | 0.9454 | 0.9336 | 0.036 | 0.952 |
| int8 | 309 MB | 0.9406 | 0.9437 | 0.9376 | 0.037 | 0.953 |
| **int8_int4_embeddings** | **218 MB** | **0.9410** | 0.9465 | 0.9356 | 0.035 | 0.953 |

**Recommended: `onnx/int8_int4_embeddings/model.onnx`** — ~5.6× smaller than FP32 at zero quality loss.

## Benchmark — Patronus Validation Set

Held-out validation split of the internal Patronus prompt-injection dataset
(5,000 sampled: 1,727 attack / 3,273 benign, max_length 512). In-distribution; quantization lossless.

| Variant | F1 | Precision | Recall | FPR | Accuracy |
|---|---|---|---|---|---|
| fp32 (parent) | 0.9971 | 0.9965 | 0.9977 | 0.002 | 0.998 |
| int8 | 0.9974 | 0.9971 | 0.9977 | 0.002 | 0.998 |
| **int8_int4_embeddings** | 0.9960 | 0.9948 | 0.9971 | 0.003 | 0.997 |

## Files
- `onnx/fp16/model.onnx`, `onnx/int8/model.onnx`, `onnx/int8_int4_embeddings/model.onnx`
- `config.json`, `tokenizer.json`, `tokenizer_config.json` (from the parent)
- `metrics/quant_bench.json`

## Usage
```python
import onnxruntime as ort, numpy as np
from transformers import AutoTokenizer
tok = AutoTokenizer.from_pretrained("patronus-studio/wolf-defender-prompt-injection-edge")
sess = ort.InferenceSession("onnx/int8_int4_embeddings/model.onnx", providers=["CPUExecutionProvider"])
enc = tok("your text", truncation=True, max_length=2048, padding="longest", return_tensors="np")
logits = sess.run(["logits"], {"input_ids": enc["input_ids"].astype(np.int64),
                               "attention_mask": enc["attention_mask"].astype(np.int64)})[0]
pred = int(logits.argmax(1)[0])  # 1 = injection
```

## 🛡️ Patronus Protect

Brought to you by **[Patronus Protect](https://patronus.studio)** — a local AI firewall that
secures every AI interaction (prompts, tools, documents) before it reaches your models.
**Try it for free at [patronus.studio](https://patronus.studio).**

## License

This model is released under the [Apache License 2.0](https://www.apache.org/licenses/LICENSE-2.0).
A copy of the license is included as `LICENSE` in this repository.

## Patronus Ark

This model is built to run inside **Patronus Ark**, Patronus' open-source on-device
AI-security scanning library (L1 native rules → L2 NTDB cascade → L3 transformer).
Ark is not publicly released yet — a repository link will be added here at launch.