Orbitune A2-512 Research-NC

Orbitune Compound Hierarchical GPT continued on the combined Aria-MIDI and GigaMIDI research corpus with bounded vectorized contiguous song-local A2 windows.

This checkpoint is for research and non-commercial use only. It is not a whole-song state-carry TBPTT model.

Artifact

Field Value
Checkpoint model.pt
Architecture orbitune-compound-hierarchical-gpt-v1
Parameters 8,857,250
Final global step 220,813
Cumulative events seen 4,096,016,384
SHA-256 e5bd2080ccf084edaa33c0df9864e4d353b4fe184ed199a2ea89a1cc06324fe0
Bytes 106,635,500
Strict load PASS

Training configuration

  • Encoding: A2 bounded vectorized contiguous song-local windows
  • Batch size: 64
  • Sequence length: 512 events
  • Precision: bfloat16
  • Dropout: 0.1
  • Optimizer: fused AdamW, restored from the parent checkpoint
  • Learning rate: 3e-4
  • Weight decay: 0.01
  • Gradient clipping: 1.0
  • torch.compile: disabled

The continuation started at global step 109,500 / 448,512,000 cumulative events. Optimizer state and Torch, Python, CUDA, and sampler RNG states were restored and preserved.

Final validation

The final checkpoint was evaluated with the fixed a2-seq512-v1 protocol:

  • loss: -1.5476370453834534
  • seed: 10001
  • 4 batches of batch size 4
  • sequence length: 512
  • validation events: 98,304
  • window hash: 0d76ff62fc73b22190f14b4377d2bc8e0da48710da398ef09efaac35fcd4da4e

Integrity and lineage

Parent checkpoint SHA-256:

4be09e0869af256f8be62e5e596b72471f15f49a6abbedc613a6a3571fd5bca8

Training source revision:

8489870f81a1591515a98e58554e533fcac9d095

The full completion record is provided in run_evidence.json; machine-readable artifact metadata is in manifest.json.

Loading

Use the Orbitune source at the recorded revision. This is a full training checkpoint, including optimizer and RNG state, rather than a Transformers-format model.

from orbitune.compound_base import CompoundHierarchicalGPT

model, payload = CompoundHierarchicalGPT.load_checkpoint(
    "model.pt",
    map_location="cpu",
)
model.eval()

License and use restrictions

  • Source code: Apache-2.0 in the Orbitune repository.
  • Model checkpoint: CC-BY-NC-SA-4.0, research/non-commercial use only.
  • Training sources include Aria-MIDI and GigaMIDI; GigaMIDI-derived training is restricted to research use.
  • Commercial use is prohibited.

No training corpus files are included in this repository.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support