everbar-v1-preview-50m

Everbar V1 Preview 50M

This is the final checkpoint from the first real Everbar V1 Preview run. It is a private testing and audition artifact, not a public release.

Model details

• Parameters: 48,875,694 (recorded training architecture count)
• Training step: 45,883 — one complete pass over the frozen preview train view
• Representation: frozen pertok-v1 symbolic music IDs; this is not an audio model
• Architecture: single-stream, block discrete diffusion, hidden size 512, 10 layers, 8 heads, MLP ratio 4, context length 768 (four represented bars at cap 192)
• Controls: none. There is no genre, instrument, role, lifecycle, or LoRA conditioning.
• Training result: final training loss 0.718426; final loss EMA 0.865831.

Everbar V1 models one pitched symbolic musical stream at a time; polyphony is supported. It does not represent synchronized multitrack V2 generation.

Intended use and limitations

Use this checkpoint only for private technical testing and short symbolic-MIDI audition through the compatible Everbar sampler. It is not an autoregressive text model and model.generate() is not its semantic generation interface.

Musical quality, safety, and licensing implications of generated material have not been publicly validated or characterized. No claim of public readiness is made. The model has bounded context and does not guarantee long-range musical coherence or support arbitrary MIDI event classes.

Compatibility and provenance

• Trained Everbar commit: 4836d7314a627d0bbe1ec94c81015197be5595a4
• Hub packaging commit: 1d9939defb767b14c1203a535266c4b8de8118f0
• Frozen snapshot manifest file SHA-256: 229709599f48b9fbd69fd0529db4b6d605ae2c46b30d778c43095ce909e3d4f3
• Frozen corpus manifest SHA-256: 28b1380b2ec712470f28c27ffc0d594c94b2d810153d99f6525ba1036f4ce804
• PerTok language: pertok-v1 / c84750b5b8a18ee1d115a372725ba0095483c83b3ca0e8973efca53d27e91bf9
• Block format: ev1-preview-clean-30k-v1-block-format-v1-cap-192 / 919abe1899468a6be33ea0d136a6194ac3bc00c1b5221992732ee70fc27588c8
• Diffusion semantics: diffusion-semantics-v1 / 476e3348b0e494c30dcfede30844e51b54b1415f70fb934a3b8d9913a3e6cfd8
• Sampler semantics: sampler-semantics-v1 / 04b4634d84ae06079dcb02e21bd03b3882bd348560a771258dfb3a2f29ae9437
• Output policy: output-policy-v1 / ea56bf939f8c295bb9976fd9a553331d84334960c0ec4d4965ac652361b137c0
• Final Safetensors SHA-256: 360f6852a8d7c7b4b17b5d907c68b38b95698e32393091538d0131ea328a4fa9

The checkpoint is bound to sampler-semantics-v1. It is not relabeled as a later sampler semantics version.

The frozen EV1_PREVIEW_ONLY clean preview contains 30,013 unique canonical streams. Its eligible source lane is PDMX's official no-license-conflict subset (CC0-1.0) plus POP909 (MIT). Motherlode preserves source provenance and attribution; raw source datasets and corpus payloads are not distributed here.

Required upstream credit includes Dreamstream Everbar, BD3-LM (Apache-2.0; revision 1c3e8f43d88dfbcee5ff2aa6932a9e74b31ae1d7), MidiTok/PerTok (MIT; revision b7478f48a05af8df34ca14058ac2b7921e03afd5), Long et al., PDMX (ICASSP 2025), and Wang et al., POP909 (ISMIR 2020).

No separate model license is assigned by this private upload. Upstream data terms and required attribution remain applicable; public redistribution and license policy require separate review.

Loading

With the installed Everbar package:

from dreamstream_everbar.huggingface import EverbarModel

model = EverbarModel.from_pretrained("nova-rey/everbar-v1-preview-50m")

The bundle also carries standalone custom code for Transformers:

from transformers import AutoModel

model = AutoModel.from_pretrained(
    "nova-rey/everbar-v1-preview-50m",
    trust_remote_code=True,
)

Pin a Hub revision for reproducible use. training-metadata.json and provenance-summary.json carry the exact training and corpus bindings.