nano4M-Audio — Team (week-1)

Week-1 data preparation for nano4M-Audio, an extension of EPFL's nano4M (the educational nano version of 4M / 4M-21) that adds audio as a fifth modality alongside RGB, depth, surface normals and captions.

This dataset covers all 12 VGGSound classes assigned to the three-person team:

person	classes
1 (Hassan)	lions roaring, horse neighing, pig oinking, cow lowing
2 (Ziyad)	dog barking, cat meowing, coyote howling, elephant trumpeting
3 (Marc)	chicken clucking, duck quacking, sheep bleating, donkey/ass braying

It contains the pre-tokenized modalities consumed by nano4M's SimpleMultimodalDataset (tok_audio@256, tok_rgb@256, scene_desc). The raw WAV+JPG pairs are kept on each teammate's machine and are not redistributed here; the tokens are the canonical artifact for training.

Source attribution. Audio + image content is derived from VGGSound (Chen et al., ICASSP 2020), whose primary key is (ytid, start). Every file in this dataset is a 1.71 s, 24 kHz mono crop centered on the CLAP-best audio peak of the corresponding VGGSound clip and the CLIP-best frame of the same window. We re-distribute the raw clips here for reproducibility under VGGSound's CC-BY-4.0 terms.

Counts

12 VGGSound classes, ~4.9k clips after link rot + CLIP/CLAP filtering. Per-class breakdown (see stats_team.md for full distributions and qualitative samples):

class	train OK	test OK	total OK
dog barking	599	38	637
lions roaring	639	4	643
duck quacking	501	38	539
chicken clucking	454	37	491
cat meowing	449	41	490
sheep bleating	405	38	443
coyote howling	344	41	385
horse neighing	325	13	338
donkey/ass braying	302	33	335
pig oinking	264	12	276
elephant trumpeting	183	34	217
cow lowing	98	13	111
TOTAL	4,563	342	4,905

All clips are uniformly 41,040 samples = 1.7100 s @ 24 kHz mono PCM16 before tokenization (frames sliced by the CLAP-best 1.71 s window from a 10 s VGGSound clip).

Files

.
├── manifest_team.csv                     canonical join table (6,529 VGGSound rows; 4,905 OK)
├── stats_team.md                         per-class counts + score histograms
├── manifest_person2.csv                  legacy Person-2 manifest (kept for reference)
├── stats_person2.md                      legacy Person-2 stats
└── tokenized/
    ├── train/tok_audio@256/{stem}.npy    EnCodec tokens, shape (1, 256), int16, [0, 1023]
    ├── train/tok_rgb@256/{stem}.npy      Cosmos tokens,  shape (1, 256), int32, [0, 63999]
    ├── train/scene_desc/{stem}.json      JSON list ["<class>"] (K=1 augmentation)
    └── test/...                          same layout

{stem} is {ytid}_{start:06d} — globally unique across the 12-class team union since (ytid, start) is VGGSound's primary key.

Tokenizer details

modality	tokenizer	shape	dtype	vocab
tok_audio@256	facebook/encodec_24khz at 1.5 kbps, K=2 RVQ codebooks at 75 Hz	(1, 256)	int16	1024
tok_rgb@256	nvidia/Cosmos-0.1-Tokenizer-DI16x16 on 256×256 frames	(1, 256)	int32	64,000
scene_desc	not pre-tokenized — captions tokenized at load time by GPT-2 ([SOS] $A [EOS], vocab 50,304)	JSON list[str]	—	—

On-disk contract for nano4M

Three details that matter for compatibility with nano4M's SimpleMultimodalDataset:

1. @256 is part of the directory name. The dataloader reads from {root}/{split}/{modality_string}/{stem}{ext} and the modality string in the config is literally "tok_audio@256", not "tok_audio".
2. scene_desc/*.json is a JSON list, not a dict. The dataloader does captions[augmentation_idx]. Run with sample_from_k_augmentations=1.
3. Captions are tokenized at load time, not pre-tokenized. The dataloader instantiates GPT-2 with [SOS] $A [EOS] template (vocab 50,304) on the fly.

Provenance & limitations

• Link rot is permanent. 385 of 2,155 source rows (17.9%) failed yt-dlp because the YouTube videos are unavailable / private / region-blocked. These are listed in raw/failed.txt with reason FAIL_DL. Re-running the downloader at a later date will lose more clips, not gain any back — that is why the raw WAV+JPG pairs are re-distributed here.
• CLIP/CLAP filtering. 26 clips dropped for FAIL_CLIP (best frame's CLIP score < 0.20 vs class label) and 14 for FAIL_CLAP (best 1.71 s window's CLAP logit < 0.20 vs class label). 1 FAIL_PEAKS (silent / sub-1.71 s clip). All in failed.txt.
• Class imbalance. elephant trumpeting has ~3× fewer clips than dog barking due to lower YouTube availability.
• Single keyframe per clip. Only the CLIP-best frame within the chosen 1.71 s window is provided, not a full video.

Reproducing

Pipeline scripts and full README are in the project repo (private). Briefly: download via yt-dlp from VGGSound's official CSV → CLAP-pick best 1.71 s window → CLIP-pick best frame in that window → EnCodec tokenize audio → Cosmos tokenize image → emit manifest. The Cosmos step requires a CUDA GPU (this dataset's tok_rgb@256 was generated on an NVIDIA L40S).

Citation

If you use VGGSound (the underlying source), please cite:

@inproceedings{chen2020vggsound,
  title={VGGSound: A Large-scale Audio-Visual Dataset},
  author={Chen, Honglie and Xie, Weidi and Vedaldi, Andrea and Zisserman, Andrew},
  booktitle={ICASSP},
  year={2020}
}

For 4M / nano4M:

@inproceedings{mizrahi20234m,
  title={4M: Massively Multimodal Masked Modeling},
  author={Mizrahi, David and Bachmann, Roman and Kar, O{\u{g}}uzhan Fatih and Yeo, Teresa and Gao, Mingfei and Dehghan, Afshin and Zamir, Amir},
  booktitle={NeurIPS},
  year={2023}
}

License

Released under CC-BY-4.0, inheriting from VGGSound. You must cite VGGSound and respect YouTube's Terms of Service when using this data.