TrustedText

High-performance AI text detection using neural networks with contrastive learning

TrustedText is a robust AI text detection system that uses deep learning to distinguish between human-written and AI-generated content. Built with neural networks and contrastive learning, it provides accurate detection with support for NVIDIA GPUs, Apple Silicon, and CPU training.

Why TrustedText?

• State-of-the-art Detection: Neural architecture with contrastive learning for robust AI text identification
• Flexible Hardware Support: Automatically detects and optimizes for NVIDIA CUDA, Apple MPS, or CPU
• Easy to Use: Simple API for both training and inference

Quick Start

Installation

# Install PyTorch (choose one based on your hardware)
pip install torch --index-url https://download.pytorch.org/whl/cu128  # NVIDIA RTX 50-series
pip install torch --index-url https://download.pytorch.org/whl/cu121  # NVIDIA RTX 40-series
pip install torch  # CPU or Apple Silicon

# Install dependencies
pip install sentence-transformers scikit-learn numpy

Note: For RTX 50-series GPUs, see the GPU Setup Guide for important compatibility information.

Train Your Model

from trustedText import TrustedText

detector = TrustedText()
detector.setup()
detector.train()

Detect AI Text

from trustedText import TrustedText

detector = TrustedText()
detector.load()

label, probability = detector.predict("Your text here")
print(f"Prediction: {label} (AI probability: {probability:.3f})")

Features

Core Capabilities

• Contrastive Learning: Learns discriminative representations by pulling same-class samples together
• Hard Negative Mining: Automatically focuses on difficult examples during training
• Embeddings Cache: Speeds up repeated training sessions
• Cross-Validation: Built-in evaluation metrics and test set support
• Batch Processing: Efficient inference on multiple texts

Technical Highlights

• Embedding Model: Qwen/Qwen3-Embedding-0.6B (768 dimensions)
• Neural Architecture: 2-layer encoder with LayerNorm and dropout
• Training Strategy: Supervised contrastive loss + cross-entropy with early stopping
• Optimizer: AdamW with cosine annealing learning rate schedule

Documentation

Guide	Description
Web Interface	Run the web UI for interactive detection
GPU Setup	Configure NVIDIA GPUs and troubleshoot CUDA
Model Transfer	Train on remote GPU machines and transfer models

Training Data Setup

TrustedText requires you to provide your own training data. Organize your data as follows:

data/
├── human/              # Human-written text files
│   ├── text001.txt
│   ├── text002.txt
│   └── ...
├── ai/                 # AI-generated text files
│   ├── text001.txt
│   ├── text002.txt
│   └── ...
└── test/              # Optional test set
    ├── human/
    └── ai/

Collection Tips:

• Aim for 100+ samples per category
• Include diverse sources (news, blogs, academic, social media)
• Use UTF-8 plain text format
• One sample per file

Usage Examples

Basic Inference

from trustedText import TrustedText

detector = TrustedText()
detector.load()

# Single prediction
label, prob = detector.predict("Your text here")

Batch Processing

detector = TrustedText()
detector.load()

texts = ["Sample 1", "Sample 2", "Sample 3"]
results = [detector.predict(text) for text in texts]

for text, (label, prob) in zip(texts, results):
    print(f"{label}: {prob:.3f} - {text[:50]}...")

Custom Configuration

from trustedText import TrustedText, Config

config = Config(
    hidden_dim=512,
    dropout=0.4,
    epochs=100,
    learning_rate=1e-3,
    contrastive_weight=0.5
)

detector = TrustedText(config=config)
detector.setup()
detector.train()

Performance Benchmarks

Training time for 1000 samples:

Hardware	Training Time
NVIDIA RTX 5060(CUDA)	~30 seconds
Apple M3 Pro(MPS)	~60 seconds

Model size: 100-500 MB (varies with training data size)

Configuration

Control training behavior with environment variables:

export MAX_SEQ_LEN=320          # Maximum sequence length
export EMBED_BATCH_SIZE=4       # Batch size (adjust for GPU memory)

Project Structure

trustedText/
├── trustedText.py          # Core library (training & inference)
├── main.py                 # CLI entry point
├── app/
│   ├── api.py             # FastAPI server
│   └── index.html         # Web interface
├── data/                  # Training data (not included)
├── models/                # Trained models (auto-generated)
└── docs/                  # Documentation

Troubleshooting

CUDA Not Available

python -c "import torch; print(torch.cuda.is_available())"

If False:

pip uninstall torch
pip install torch --index-url https://download.pytorch.org/whl/cu121

Out of Memory

Reduce batch size:

export EMBED_BATCH_SIZE=2
python trustedText.py

Model Not Found

Train a model first or transfer from another machine:

python trustedText.py  # Train locally

See Model Transfer Guide for remote training.

Contributing

Contributions are welcome! Please follow these steps:

1. Fork the repository
2. Create a feature branch (git checkout -b feature/improvement)
3. Make your changes
4. Submit a pull request

See CONTRIBUTING.md for detailed guidelines.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Citation

If you use TrustedText in your research or project, please cite:

@misc{trustedtext2026,
    title={TrustedText: AI Text Detection using Neural Networks with Contrastive Learning},
    author={Timmy Wu and Matthew Hung},
    year={2026},
    publisher={GitHub},
    howpublished={\url{https://github.com/theoncetimes/trustedText}}
}

Support

• Documentation: Check the docs/ folder
• Issues: Open an issue on GitHub

---

Made with focus on accuracy, performance, and ease of use.