README.md

metadata

license: mit
library_name: pytorch
tags:
  - music-generation
  - symbolic-music
  - abc-notation
  - classical
  - transformer
pipeline_tag: text-to-audio

NotaGen-X — MAESTRO Mirror

Safetensors mirror of the NotaGen-X model weights, used by MAESTRO for the AI-Workstation "Create → NotaGen" model card.

This repository ships only the model weights (no optimizer state) in the safetensors format, so they load into MAESTRO without arbitrary-code execution risk. The original upstream checkpoints from ElectricAlexis/NotaGen are PyTorch .pth files bundling optimizer state, epoch counters, etc.

Files

File	Purpose
model.safetensors	NotaGen-X model state dict (516 M params, fp32, ~1.92 GB)
config.json	Architecture hyperparameters consumed by NotaGen's loader
README.md	This file

Provenance

• Upstream: ElectricAlexis/NotaGen · file weights_notagenx_p_size_16_p_length_1024_p_layers_20_h_size_1280.pth
• Conversion: torch.load(...)['model'] → contiguous CPU tensors → safetensors.torch.save_file(...). Round-trip verified (zero mismatches over all 323 tensors).
• Training metadata preserved in safetensors header (epoch, best_epoch, min_eval_loss).

Architecture

Hierarchical Tunesformer (GPT-2 backbone) generating ABC notation:

• p_size = 16 (patch size)
• p_length = 1024 (patch sequence length)
• p_layers = 20 (patch-level decoder layers)
• c_layers = 6 (character-level decoder layers)
• h_size = 1280 (hidden size)

Loading

from safetensors.torch import load_file
state = load_file("model.safetensors")
# Pass to NotaGen's model class — same key layout as the upstream `.pth`'s
# `["model"]` sub-dict.

MAESTRO's loader (backend/ai/models/notagen.py) handles this automatically and, if the upstream NotaGen inference script demands a .pth on disk, materialises one alongside the safetensors on first load.

License

MIT, inherited from upstream NotaGen.