The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.
Error code: FileFormatMismatchBetweenSplitsError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
v2a-vjepa — research artifacts
Artifacts of the research project "V-JEPA 2.1 features as additional conditioning for MMAudio-L video-to-audio" (VGGSound, dev split 768 clips). Code and full documentation: https://github.com/jeongyh98/v2a-vjepa (branch recover-0916) — see docs/RESULTS_0923.md (results §1–§19), docs/PIPELINE_V9_0926.md, docs/DECISIONS_LOG.md, docs/HANDOFF.md.
Headline (as of 2026-10-02): every attempt to make the generator use V-JEPA content (read paths, relation modules, CLAP alignment V9, oracle-acoustic conditioning V10-E0, decoder adaptation) failed its pre-registered success criteria. The one robust effect — FD-PaSST −8.5 vs public MMAudio-L from a frozen generator + 5.36M read path — is reproduced with mismatched (wrong-video) features and is therefore not attributable to video content (§12). Human listening on the blind packs is still pending. Treat everything here as negative/diagnostic results, not a method release.
What is here (paths mirror results/ in the repo)
| path | content |
|---|---|
*/best.pth (55 files, 2.3 GB) |
trained read paths (V6/V7/V8), alignment adapters (V9 align_*), generation-stage readers with adapters (g_*, fz_*, ad_*), V10 E0 acoustic read path (v10/e0_s0) |
*/gen_dev_*/*.flac (~27 GB) |
generated dev-768 audio for every arm, 8 s / 44.1 kHz, Euler 25, CFG 4.5, noise seed 42 (per-arm; _other = wrong-clip conditioning). exp1/gen_dev_pub = public MMAudio-L |
r0_night/r0n_20260928_0259/ |
R0 candidate bank: 640 clips × 4 candidates (per-candidate SHA-256 seeds in manifest.json), QC, per-candidate IB/DeSync/onset-F1, R0b training-free selection stats, REPORT.md. Listening pack without videos and without keys |
v9/clap_c20, v9/clap_dev |
frozen laion-CLAP (HTSAT-base) 2 s segment embeddings (+ log-mel reference, dB) |
v10/atst_c20, v10/atst_dev |
frozen ATST-Frame (base) 0.25 s × 32 and 40 ms frame features (last block, 768-d) |
v8/lv_train, v8/lv_dev |
window-pooled (W=4) V-JEPA layer variants + sync/clip/audio-latent targets used by readout probes |
exp1/dev_gt_cache |
av-benchmark feature caches for the dev-768 ground truth (audio features + ImageBind/Synchformer video features) |
**/avb_dev_*/output_metrics.json, **/judge_dev_*.json, */*stats*.json |
av-benchmark metrics, onset judge per-clip results, paired-bootstrap statistics |
v10/behav128/, v9/listen/ |
blind listening packs (A/B FLAC only) |
Not included: VGGSound videos (no redistribution rights), ground-truth audio waveforms, the raw V-JEPA 2.1 token caches (≈500 GB; reproducible with probe/v8/build_shortlong.py from the official Meta checkpoint vjepa2_1_vitl_dist_vitG_384.pt, verified bit-identical to the HF port — §18.4), third-party weights (MMAudio, V-JEPA 2.1, CLAP, ATST-Frame — use their official sources), and listening-pack answer keys (private_keys/, KEY_do_not_open.json) which stay offline until human evaluation is finished.
Provenance / license
Generated audio and features are derived from VGGSound (YouTube) clips conditioned through MMAudio-L (large_44k_v2), V-JEPA 2.1, laion-CLAP and ATST-Frame; each upstream has its own license and the VGGSound clips themselves are not redistributed. These artifacts are shared for research reproducibility only. The test split (15,341) was never used for generation, scoring, or model selection.
Reproduce
git clone https://github.com/jeongyh98/v2a-vjepa && git checkout recover-0916, place this dataset under results/, follow docs/HANDOFF.md. Noise seeds, splits and all pre-registered protocols are in the docs.
- Downloads last month
- 78