arpanpathak/staleynet-data

StaleyNET derived training data

Derived artifacts for training and evaluating StaleyNET, a text-prompted music generator written in Rust. This repository contains the caches the training loop reads, the fitted RVQ tokenizer, generated samples, and the metadata needed to rebuild the caches from the original datasets. It does not contain the raw third-party audio.

Code, paper, and book: https://github.com/arpanpathak/staleynet. Checkpoints: https://huggingface.co/arpanpathak/staleynet-checkpoints.

Release v0.1.2

• GitHub release: https://github.com/arpanpathak/staleynet/releases/tag/v0.1.2
• staleynet-audio 0.1.2 on crates.io: https://crates.io/crates/staleynet-audio. Its page explains each audio step with intuition and runnable examples.
• Since v0.1.1: the GPU kernels are organised into one module per topic, the AGX Thor builds against cuda-oxide's main branch, and the paper has been revised with hardware, benchmark and tokenizer appendices.
• New in these repositories: the lyrics-thor, lyrics-thor-long and pq-large runs, and the band-group token caches they train on.
• CI: rustfmt, clippy with warnings denied, cargo-deny, the audio coverage gate, and documentation tests for every example on the crate page.

At a glance

Item	Value
Files	3,189
Size	3.12 GB
Audio features	128 mel bands at 24 kHz, FFT 1024, hop 256
Audio tokens	4 residual codebooks, 1024 entries each
Tokenizer	residual k-means, about 12 MB
Samples	79 generated, reconstruction, tokenizer, and studio files
License	CC BY-SA 4.0, subject to the upstream dataset terms

Listening in the dataset viewer

The viewer above plays the project's audio outputs, not its training data. The training data itself is the spectrogram and token caches described below, binary files that the viewer cannot play. Choose a view from the menu:

View	What it contains
reconstruction (default)	Four held-out songs, each in four versions: the original recording, its exact spectrogram through Griffin-Lim, its tokens using only the coarsest level, and its tokens using all 16 codes. This is the clearest way to hear where detail is lost; the spectral_error_log_units column gives the measured loss of each version.
reconstruction-first-tokenizer	The same comparison with the first tokenizer, which loses about twice as much detail.
generated	Samples generated by the early models on the first tokenizer. They sound blurred and noisy: the reconstruction views show that most of that blur comes from the first tokenizer, which the current one halves.
overfit-test	A model trained on a single excerpt reproduces it; compare its outputs with the excerpt's own tokens.
recordings	The author's own takes, with their descriptions and lyrics. All rights reserved.

Layout

Directory	Files	Size	Contents
cache/	1,415	2.0 GB	SDD log-mel (.f16) and RVQ codes (.u16), tokenizer, manifest
cache-jamendolyrics/	161	408 MB	JamendoLyrics log-mel, codes, tokenizer, manifest
cache-g4l4f1/	708	246 MB	band-group tokenizer (4 groups × 4 levels): SDD codes, tokenizer, manifest
cache-pq-jamendolyrics/	82	56 MB	JamendoLyrics codes under the band-group tokenizer
cache-pq-recordings/	5	6.1 MB	codes for the author's own takes under the band-group tokenizer
cache-l16f1/	709	264 MB	tokenizer-study control: 16 residual levels without band groups
cache-overfit/	5	12 MB	single-take overfit cache
cache-recordings/	3	2.2 MB	log-mel for the author's own takes
recordings/	8	8.0 MB	2 own takes: WAV, caption, lyrics, JSON
overfit/	2	572 KB	the single overfit take and caption
samples/	79	34 MB	generated, reconstruction, tokenizer, and studio samples
raw/	7	480 KB	SDD and JamendoLyrics metadata and license tables, plus the provenance README

The caches built on another tokenizer reuse the log-mel features of their source cache, so their mel/ directories are omitted here to avoid duplicating several gigabytes. Recreate them with:

ln -s ../cache/mel data/cache-g4l4f1/mel
ln -s ../cache/mel data/cache-l16f1/mel
ln -s ../cache-jamendolyrics/mel data/cache-pq-jamendolyrics/mel
ln -s ../cache-recordings/mel data/cache-pq-recordings/mel

The band-group caches feed the pq-large run in the checkpoint repository. cache-l16f1/ is the control of the tokenizer study: its held-out error of 0.427 against 0.401 for cache-g4l4f1/ shows that most of the improvement over the original tokenizer comes from using more codes per step.

Formats

• mel/*.f16: little-endian f16 log-mel, 128 bands, 24 kHz, FFT 1024, hop 256, 0 to 12 kHz, normalized with the per-band statistics in manifest.json.
• codes/*.u16: little-endian u16 RVQ indices, 4 codebooks of 1024 entries, 1 token per codebook per audio step.
• tokenizer.json: the fitted residual k-means codebooks, about 12 MB.
• manifest.json: track list, mel parameters, normalization statistics, and per-track token counts.
• samples/*.json: the prompt, checkpoint, timing, and real-time factor next to each generated WAV.

Download and use

hf download arpanpathak/staleynet-data \
  --repo-type dataset --local-dir data

# The caches now sit where the CLI expects them.
staleynet sing  --cache data/cache --run ... --prompt "..." --out song.wav
staleynet train --cache data/cache --run runs/base --config configs/base.json

For the lyrics model, add data/cache-jamendolyrics through extra_sources in the run configuration, as configs/lyrics-orin.json does.

Provenance

• Song Describer Dataset (SDD), https://zenodo.org/records/10072001. The dataset has 706 Jamendo recordings with human captions. Captions are CC BY-SA 4.0. Each recording carries the Creative Commons license listed in raw/song-describer/audio_licenses.txt, including CC BY-NC-ND and CC BY-NC-SA. Raw audio is not redistributed here.
• JamendoLyrics MultiLang, https://github.com/f90/jamendolyrics. Code and metadata are MIT. Songs and lyrics carry the per-song CC license in raw/jamendolyrics/JamendoLyrics.csv. The upstream repository notes that the dataset has moved to https://huggingface.co/datasets/jamendolyrics/jamendolyrics. Raw audio is not redistributed here.

To rebuild the caches, follow the commands in the code README:

staleynet prepare  --song-describer data/raw/song-describer --cache data/cache
staleynet tokenize --cache data/cache
staleynet prepare  --jamendo-lyrics data/raw/jamendolyrics --cache data/cache-jamendolyrics
staleynet tokenize --cache data/cache-jamendolyrics --reuse data/cache/tokenizer.json

Licensing

The derived caches, tokenizer, and generated samples are offered under CC BY-SA 4.0 (LICENSE-CC-BY-SA-4.0.txt) only to the extent the upstream dataset licenses permit it. The source recordings include NonCommercial and NoDerivatives tracks, so do not use these artifacts commercially where the upstream license forbids it.

Carve-out: recordings/, overfit/take.wav, and cache-recordings/ are the author's own vocal recordings. They are included for reproducibility and are not licensed for reuse. All rights are reserved. Do not redistribute them or train a voice model on them without permission.

raw/ contains only metadata tables copied from the upstream datasets. Their own licenses apply.

Integrity

SHA256SUMS lists every file in this repository. Verify with:

sha256sum -c SHA256SUMS

Citation

@misc{pathak2026staleynet,
  author = {Arpan Pathak},
  title  = {StaleyNET: Text-Conditioned Music Generation Trained from Scratch
            on Edge GPUs in Rust},
  year   = {2026},
  url    = {https://github.com/arpanpathak/staleynet}
}