SuperAnnotate AI Detector
| Property | Value |
|---|---|
| Base Model | RoBERTa Large |
| License | SAIPL |
| Language | English |
| Repository | Available on GitHub |
What is ai-detector?
The SuperAnnotate AI Detector is a sophisticated model designed to distinguish between human-written and AI-generated text. Built on RoBERTa Large architecture, it achieves remarkable accuracy across various LLM outputs, with particularly high performance on modern models like ChatGPT (99.2% accuracy) and GPT-4 (98.5% accuracy).
Implementation Details
The model utilizes a custom binary classification architecture, trained on a carefully curated dataset of 44,000 text samples. The training process employed specific strategies including label smoothing (0.38), SGD optimizer, and cosine scheduling with hard restarts. The model was trained for 20 epochs with a learning rate of 5e-05 and gradient clipping at 3.0.
- Balanced dataset structure with equal human/AI content representation
- Training data sourced from Wikipedia, Reddit ELI5 QA, and Scientific Papers
- Implements advanced n-gram filtering to prevent overfitting
- Tested against 11 different LLM models and 11 adversarial attacks
Core Capabilities
- High accuracy detection of text from modern LLMs (85.2% average accuracy)
- Robust performance across multiple domains and text types
- Effective against various adversarial attacks
- Calibrated confidence scores for reliable predictions
Frequently Asked Questions
Q: What makes this model unique?
The model stands out for its balanced training approach and high accuracy across modern LLMs, particularly achieving over 99% accuracy on ChatGPT outputs while maintaining good performance on human text detection (73.1% accuracy).
Q: What are the recommended use cases?
The model is particularly suited for educational institutions detecting AI-generated assignments, research organizations validating authentic content, and content platforms seeking to identify synthetic text. It's also valuable for training data validation and fraud detection in scientific writing.
