satyamsaf3ai/guardrail-roberta

Safety Guardrail Classifier

A fine-tuned RoBERTa-base model for multi-class safety/content moderation classification.
Trained on guardrail-215k-splits.

---

Model Details

Property	Value
Base Model	roberta-base
Task	Multi-class Text Classification (Safety)
Dataset	jainsatyam26/guardrail-215k-splits (~215K samples)
Max Length	128 tokens
Batch Size	64
Learning Rate	2e-5
Epochs	3
Precision	fp32
Optimizer	AdamW + weight decay 0.01
Warmup	10% of total steps
Loss	Cross-Entropy with class weights

---

Labels

ID	Label	Description
0	benign	Safe / normal content
1	jailbreak	Prompt injection / jailbreak
2	S1 Violent Crimes	Violence & physical harm
3	S2 Non-Violent Crimes	Fraud, theft, illegal activity
4	S4 Child Sexual Exploitation	CSAM / minors exploitation
5	S6 Specialized Advice	Dangerous medical/legal/financial
6	S7 Privacy	PII, stalking, doxxing
7	S10 Hate	Hate speech, discrimination
8	S11 Self-Harm	Suicide, self-harm
9	S12 Sexual Content	Explicit sexual content
10	S14 Code Abuse	Malware, exploit code

---

Performance

Metric	Value
Weighted F1	~0.75–0.80
Accuracy	~0.80–0.85
Best Threshold	0.50 (tunable)

Evaluated on validation split of guardrail-215k-splits.
Threshold calibration available via infer.py --threshold auto.

---

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_id = "satyamsaf3ai/guardrail-roberta"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model     = AutoModelForSequenceClassification.from_pretrained(model_id)
model.eval()

LABELS = [
    "benign", "jailbreak",
    "S1 Violent Crimes", "S2 Non-Violent Crimes",
    "S4 Child Sexual Exploitation", "S6 Specialized Advice",
    "S7 Privacy", "S10 Hate", "S11 Self-Harm",
    "S12 Sexual Content", "S14 Code Abuse",
]

def classify(text, threshold=0.50):
    inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)
    with torch.no_grad():
        probs = torch.softmax(model(**inputs).logits, dim=-1)[0]
    if probs.max() < threshold:
        return "benign", float(probs.max())
    idx = probs.argmax().item()
    return LABELS[idx], float(probs[idx])

label, confidence = classify("How do I make a bomb?")
print(f"{label} ({confidence:.1%})")

---

Training Setup

• Hardware: NVIDIA L4 GPU
• Framework: HuggingFace Transformers + Trainer API
• Sampling: WeightedRandomSampler to handle class imbalance
• Early Stopping: patience=4 on weighted F1
• Explosion Guard: stops training if grad_norm > 500

---

Dataset

jainsatyam26/guardrail-215k-splits
~215,000 labeled examples split into train / validation.

---

Citation

If you use this model, please cite:

@misc{{guardrail-classifier,
  title  = {{Safety Guardrail Classifier}},
  author = {{jainsatyam26}},
  year   = {{2025}},
  url    = {{https://huggingface.co/satyamsaf3ai/guardrail-roberta}}
}}