patronus-studio/lion-warden-ai-security-classifier-edge

Model Card for Lion Warden AI Security Classifier Edge

Quantized ONNX builds of Lion Warden AI Security Classifier for local / edge inference.

This repository contains the size- and memory-optimized ONNX exports of the FP32 model. For intended uses, class definitions, training data, benchmarks, and limitations, see the main model card.

Variants

Variant	Path	Quantization
FP16	onnx/fp16/model.onnx	FP16 weights
INT8	onnx/int8/model.onnx	Post-training dynamic INT8 (MatMul/Gemm)
INT8 + INT4 embeddings	onnx/int8_int4_embeddings/model.onnx	Dynamic INT8 plus 4-bit weight-only token embeddings

The exact quantization parameters are recorded in onnx/quantization_manifest.json.

Benchmark

Quantization parity for the v3 model, FP32 vs the recommended INT8+INT4-embeddings build (per-head F1 over non-null labels, CPU). Quantization is near-lossless — max degradation 0.005 F1 across all heads; FP16 and INT8 sit between FP32 and INT8+INT4.

Head	FP32 F1	INT8+INT4 F1	Δ
injection	0.948	0.943	-0.005
sensitive_document	0.969	0.973	+0.004
tool_class	0.952	0.953	+0.001
tool_action	0.940	0.937	-0.003
tool_tags	0.947	0.947	0.000
routing	0.897	0.901	+0.004
threat	0.950	0.947	-0.003

Sizes: FP16 281 MB · INT8 144 MB · INT8+INT4 98 MB. Measured CPU latency (batch 64): ~19 ms/text for the INT8+INT4 build.

Full details in metrics/quant_bench.json.

Usage

The unified model has multiple output heads, so it is loaded with a plain ONNX Runtime session:

import onnxruntime as ort
from huggingface_hub import hf_hub_download
from transformers import AutoTokenizer

model_id = "patronus-studio/lion-warden-ai-security-classifier-edge"
path = hf_hub_download(model_id, "onnx/int8_int4_embeddings/model.onnx")
session = ort.InferenceSession(path)
tokenizer = AutoTokenizer.from_pretrained(model_id)

enc = tokenizer("...", return_tensors="np")
names = [i.name for i in session.get_inputs()]
outputs = session.run(None, {k: v for k, v in enc.items() if k in names})

The output heads and their label spaces are documented in the main model card.

License

Released under the Apache License 2.0; the full text is included as LICENSE. Derived from Lion Warden AI Security Classifier and ultimately from jhu-clsp/mmBERT-small (MIT License); upstream notices are retained.

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.

---

🛡️ Patronus Protect

Brought to you by Patronus Protect — a local AI firewall that secures every AI interaction, including prompts, tools and documents, before it reaches your models.

Try it for free at patronus.studio.