Architecture: DeBERTa-v3-base
Training Data: Custom dataset of Human Wikipedia articles vs. AI-generated Wikipedia-style articles (GPT-Neo).
Performance: Achieved 99.4% Accuracy on in-distribution test data.

Limitations (The "Generalization Gap")

This model was trained to detect GPT-Neo (1.3B).

Performance on GPT-Neo: 99.9% Confidence

This highlights the necessity of domain-specific training for AI detection. Detectors trained on legacy models (2021 era) cannot reliably detect modern SOTA models (2024 era) due to the closing gap in perplexity and burstiness.

How to Use

python

from transformers import pipeline

classifier = pipeline("text-classification", model="vraj33/ai-text-detector-deberta")
text = "The quick brown fox..."
result = classifier(text)
print(result)