AURORA-Transformer (ART) Implementation

A modular reasoning architecture for reliable and efficient neural systems under extreme low-precision (NVFP4) computation.

Project Structure

art/
├── art/                    # Core library
│   ├── modules/           # ART components
│   │   ├── bpl.py        # Binary Perception Layer
│   │   ├── ctc.py        # Cognitive Transformer Core
│   │   ├── d3r.py        # Dynamic DAG Router
│   │   ├── cra.py        # Context & Reality Anchor
│   │   ├── syn.py        # Synthesis Head
│   │   └── art.py        # Full ART model
│   ├── training/         # Training utilities
│   │   ├── trainer.py    # Training loop
│   │   ├── dataset.py    # Data loading
│   │   └── metrics.py    # Evaluation metrics
│   └── utils/            # Utilities
│       ├── config.py     # Configuration
│       └── quantization.py # NVFP4 simulation
├── api/                   # Deployment
│   └── server.py         # FastAPI server
├── configs/              # Configuration files
├── data/                 # Training data
├── checkpoints/          # Model checkpoints
├── scripts/              # Training scripts
└── tests/                # Unit tests

Installation

pip install -e .

Training

python scripts/train.py --config configs/art_base.yaml

Inference

python scripts/inference.py --checkpoint checkpoints/best.pt --input "Your query"

API Server

uvicorn api.server:app --host 0.0.0.0 --port 8000

Citation

@article{biswas2026art,
  title={AURORA-Transformer (ART): A Modular Reasoning Architecture},
  author={Biswas, Swadhin},
  year={2026}
}