mean list | std list | frames int64 | n_files int64 | manifest list |
|---|---|---|---|---|
[
0.2544849917268572,
-0.12198937206916516,
-0.010053251759337297,
-0.08641316137192688,
0.0636601761768548,
-0.0667114214316907,
0.012210764924615973,
0.028573196351109203,
-0.08323610942553004,
-0.041251268261996626,
-0.038016536048696786,
-0.023131808138696246,
0.07445111874032852,
0.0517... | [
1.0411281461334894,
0.9560335885691397,
0.9160925208493558,
0.8834822819338377,
0.9466352538916213,
0.9272088692513497,
0.8832150546124647,
0.8720315751730594,
0.933516222050337,
0.9264367396076914,
0.9458185280229375,
0.9251713196653699,
0.9219739354617876,
0.9799504951872332,
0.9086770... | 155,070 | 74 | [
"data/proc/test/instr_181523.npz",
"data/proc/test/instr_2076.npz",
"data/proc/test/instr_2191.npz",
"data/proc/test/instr_2218.npz",
"data/proc/test/instr_2243.npz",
"data/proc/test/instr_2304.npz",
"data/proc/test/instr_2382.npz",
"data/proc/test/instr_2477.npz",
"data/proc/test/instr_2575.npz",
... |
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
singing-fm data/proc — YuE2-VAE latents + MERT-v2 features
Preprocessed training data for the singing-fm latent flow-matching project. Derived features only — no raw audio.
Contents
data_proc.tar.zst (tar.zst, ~6.5 GB) extracts to data/proc/:
| Path | What |
|---|---|
train/*.z.npy |
1462 files, fp16, YuE2-VAE posterior-mean latents z [T, 64] @ 25 Hz |
train/*.h.npy |
1462 files, fp16, MERT-v2 (layer 21) frame features h [T, 1024] @ 25 Hz (aligned to z) |
test/*.npz |
74 holdout files with z + h (KAD reference / cond source) |
stats.json |
per-channel z mean/std used for train-time normalization (computed over test npz — replicate this exactly when de-normalizing) |
stats.json is also at repo root for inspection.
Domains (filename prefixes)
| Prefix | Source | Train files |
|---|---|---|
speech_ |
EARS (Meta) | 540 |
singing_ |
VocalSet | 531 |
song_ |
Song Describer Dataset (vocal) | 180 |
instr_ |
MusicNet (165) + SDD instrumental (46) | 211 |
Test split = every 20th file of each domain's sorted list (--holdout_every 20),
74 files total, 1.7 h.
Reproduction
Built by the pipeline in the singing-fm repo (preproc.py): audio → YuE2-VAE
(FP32) latents + MERT-v2 (FP16 weights, outside autocast) layer-21 features,
both truncated to the common frame count. Deterministic given the source
datasets (sorted globs, fixed holdout, random.seed(0) for MusicNet selection).
Sources & licenses
Features derived from four public research datasets: EARS, VocalSet, Song Describer Dataset, MusicNet. No raw audio is included; consult the original dataset releases for their respective licenses and citation requirements.
- Downloads last month
- 40