🐾 PurrBERT-v1.1

PurrBERT-v1.1 is a lightweight content-safety classifier built on top of DistilBERT.
It’s designed to flag harmful or unsafe user prompts before they reach an AI assistant.

This model is trained on a combination of:

• HateCheck
• Toxic Completions
• Aegis AI Content Safety Dataset 2.0

---

📝 Model Description

• Architecture: DistilBERT with a classification head (2 labels: SAFE vs. FLAGGED)
• Purpose: Detect hate speech, toxic content, and unsafe prompts in English text.
• Input: A single string (prompt text).
• Output: A binary prediction:
  • 0 → SAFE
  • 1 → FLAGGED

---

🧠 Training Details

• Base model: distilbert-base-uncased
• Epochs: 2 (initial run)
• Optimizer: AdamW
• Batch size: 16
• Learning rate: 2e-5
• Weight decay: 0.01

Loss dropped steadily during training, and metrics were evaluated on a held-out test set.

---

📊 Evaluation Results

On an Aegis test slice:

Metric	Score	v1	v1.1
Accuracy		0.8050	0.8200
Precision		0.7731	0.8091
Recall		0.8846	0.8558
F1 Score		0.8251	0.8318

Latency per prompt on GPU: ~0.0230 sec

---

🚀 Usage

from transformers import DistilBertTokenizerFast, DistilBertForSequenceClassification
import torch

# Load trained model and tokenizer
model = DistilBertForSequenceClassification.from_pretrained("purrgpt-community/purrbert-v1.1")
tokenizer = DistilBertTokenizerFast.from_pretrained("purrgpt-community/purrbert-v1.1")
model.eval()

def classify_prompt(prompt):
    inputs = tokenizer(prompt, return_tensors="pt", truncation=True, padding=True)
    with torch.no_grad():
        outputs = model(**inputs)
        pred = torch.argmax(outputs.logits, dim=-1).item()
    return "SAFE" if pred == 0 else "FLAGGED"

print(classify_prompt("You are worthless and nobody likes you!"))
# → FLAGGED

---

⚠️ Limitations & Bias

• The model is trained primarily on English datasets.
• It may produce false positives on edgy but non-harmful speech, or false negatives on subtle harms.
• It reflects biases present in its training datasets.

---

🐾 Intended Use

PurrBERT is intended for moderating prompts before they’re passed to AI models or for content-safety tasks. It is not a replacement for professional moderation in high-risk settings.