itsVentie/nano-guard-classifier-v1 is a binary sequence classification model fine-tuned for real-time prompt injection and jailbreak attack identification in LLM pipelines.

Source Code & Engine Integration: itsVentie/nano-guard on GitHub

Model Details

Base Architecture: answerdotai/ModernBERT-base
Task: Binary Sequence Classification (0: SAFE, 1: ATTACK)
Max Sequence Length: 256 tokens
Optimization Objective: Low-latency CPU inference via ONNX Runtime / ort.

Evaluation Results

Evaluated on test splits of itsVentie/llm-red-teaming-corpus and itsVentie/rag-prompt-injection-bench.

Metric	Value
Accuracy	0.982
Precision	0.978
Recall	0.985
F1 Score	0.981

Usage

PyTorch / Transformers

python

import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification

model_id = "itsVentie/nano-guard-classifier-v1"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)

inputs = tokenizer("System prompt override instruction", return_tensors="pt", truncation=True, max_length=256)
with torch.no_grad():
    outputs = model(**inputs)
    probabilities = torch.softmax(outputs.logits, dim=-1)

risk_score = probabilities[0][1].item()
print(f"Attack Probability: {risk_score:.4f}")