achrafness/SafeGuardX-v2

SafeGuard — SEES Blue Team

Labels

ID	Label	Meaning
0	SAFE	No violation
1	VIOLATION	Prompt injection / policy violation

Challenge I/O

Input: {"conversation": [{"role": "...", "content": "..."}]} Output: {"violation": true, "confidence": 0.97}

Usage

import json, torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification

tokenizer = AutoTokenizer.from_pretrained("achrafness/SafeGuardX-v2")
model = AutoModelForSequenceClassification.from_pretrained("achrafness/SafeGuardX-v2")
model.eval()

def classify(conversation):
    text = json.dumps(conversation)
    inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=512)
    with torch.no_grad():
        logits = model(**inputs).logits
    conf = float(torch.softmax(logits, dim=-1)[0][1])
    return {"violation": conf > 0.5, "confidence": round(conf, 4)}