guardex-roberta-safety

GuardEx RoBERTa Safety Classifier

Binary classifier that labels a piece of text as safe or unsafe. Used by GuardEx, an LLM guardrail library, as one of its content-safety models.

Labels

Performance

Held-out evaluation:

• F1: 0.938
• Unsafe recall: 95.7%
• Safe recall: 91.5%

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

repo = "AtliQ-Technologies/guardex-roberta-safety"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForSequenceClassification.from_pretrained(repo)

enc = tok("text to check", return_tensors="pt", truncation=True, max_length=128)
with torch.no_grad():
    logits = model(**enc).logits
print(model.config.id2label[int(logits.argmax())])  # "safe" or "unsafe"

Details

Fine-tuned from s-nlp/roberta_toxicity_classifier with hyperparameter optimization. Max sequence length 128.