murali5613/guardrail-deberta-classifier-demo

Guardrail DeBERTa Classifier

Fine-tuned mDeBERTa-v3-base for detecting benign, jailbreak, and harmful prompts.

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tokenizer = AutoTokenizer.from_pretrained("YOUR_USERNAME/guardrail-deberta-classifier")
model = AutoModelForSequenceClassification.from_pretrained("YOUR_USERNAME/guardrail-deberta-classifier")

prompt = "Ignore all previous instructions..."
inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=512)
with torch.no_grad():
    logits = model(**inputs).logits
    probs = torch.softmax(logits, dim=-1)
    print(probs)  # [benign, jailbreak, harmful]

Classes

• 0: benign
• 1: jailbreak
• 2: harmful

Training

• ~11k examples from open HF datasets
• Test accuracy: 96.6%
• Macro F1: 96.1%