AuraMIDI-v1 is a lightweight, high-performance symbolic music generation Transformer model trained using PyTorch and exported to ONNX for universal deployment across web, mobile, and desktop environments.

Trained on a curated dataset of 19,833 high-quality songs (~80 Million musical tokens) from the Lakh Clean MIDI Dataset, Google Magenta MAESTRO, and modern multi-genre stems (Synthwave, Cyberpunk, Lo-Fi Chill, Neo Jazz, Phonk, and EDM).

📊 Model Architecture & Training Specs

Model Name: AuraMIDI-v1
Model Type: Causal Decoder-Only Transformer (GPT-style Music Architecture)
Total Parameters: 48.60 Million
Transformer Layers: 12 Layers (Pre-LN with GELU activation)
Attention Heads: 16 Heads ($d_{head} = 32$)
Hidden Size ($d_{model}$): 512
Feedforward Dim: 2048
Context Length: 1024 tokens
Vocabulary Size: 10,000 (MidiTok REMI Tokenizer)
Training Dataset: 19,833 songs / 79,454,977 tokens
Training Hardware: NVIDIA Modern GPU with Mixed Precision (BF16 AMP)
Training Epochs: 3 full epochs (~238 Million tokens processed)
Final Validation Loss: 1.4065

📦 Files in this Repository

File	Size	Description
`auramidi_v1.pt`	~194 MB	Full PyTorch checkpoint (Loss 1.40)
`auramidi_v1.onnx`	~166 MB	Monolithic standalone ONNX model for Web/Mobile
`generate_neural.py`	—	Autoregressive generation script with Top-K/Top-P sampling
`model_architecture.py`	—	PyTorch model architecture and tokenization pipeline

🚀 Quick Start (Inference in Python)

pip install torch miditok pretty_midi

python

import torch
from generate_neural import generate_full_neural_composition

# Generate a complete 4-track neural track
midi_path = generate_full_neural_composition(
    genre="synthwave",
    key="Am",
    bpm=120,
    temperature=0.85,
    top_k=20,
    top_p=0.90
)

print(f"Generated neural MIDI: {midi_path}")

nitrai-research/AuraMIDI-v1

⚡ AuraMIDI-v1 - Neural Music Transformer

📊 Model Architecture & Training Specs

📦 Files in this Repository

🚀 Quick Start (Inference in Python)

nitrai-research/
AuraMIDI-v1