HowTo1k — a 128-hour HowTo100M subset for video SSL pretraining

A reproduction of the "0.1% of HowTo100M = 128 hours" data-scale cell used in Intuitive physics understanding emerges from self-supervised pretraining on natural videos (Garrido, Ballas, LeCun), preprocessed for V-JEPA-style training.

Videos	1,119 (whole videos, not caption clips)
Total duration	128.01 h
Mean / median length	412 s / 312 s
Resolution	short side ≥ 224 (median 360), aspect ratio preserved
Frame rate	30 fps (all files)
Codec	H.264, CRF 23
Size	22 GB

How it was built

1. Video list — csv/howto100m_videos.csv from antoine77340/MIL-NCE_HowTo100M, the authors' mirror manifest: 1,238,791 entries.
2. Sampling — uniform draw at a fixed seed (shuf --random-source, seed 0, plus a 300-entry top-up at seed 1 from the complement). This follows the paper's protocol, "sample uniformly X% of the videos" — whole videos, not the ASR caption segments that HowTo100M calls "clips". Those exist for text-video alignment; a vision-only model never reads them, and indexing them would change the sampling distribution from per-video-uniform to per-second-uniform. sampled.txt records the full draw in order.
3. Download — yt-dlp, best stream with height<=360.
4. Transcode — short side capped at 360 with min() so nothing is upsampled, aspect ratio preserved, 30 fps, H.264. No square centre crop: V-JEPA does its own RandomResizedCrop(scale=(0.3,1.0), ratio=(0.75,1.35)), and pre-squaring would discard 44% of a 16:9 frame first.
5. Filter — short side ≥ 224 and ≥ 96 frames (num_frames 16 × sampling_rate 6).

Why short side 360

With RandomResizedCrop(scale=(0.3,1.0), ratio=(0.75,1.35)) targeting 224×224, the fraction of crops that end up upsampled depends on the source's short side (measured, 200k samples):

source short side	224	256	288	320	360
crops upsampled	98.3%	64.5%	33.5%	8.4%	0.0%

At 224 the crop height can never exceed the source height, so almost every sample is interpolated. 360 removes this; going higher only costs disk and decode time.

Known deviations from the paper

• 1,119 videos, not 1,239. 0.1% of 1,238,791 is 1,239. The hours match (128.01 vs 128) because the videos drawn here average 412 s against the ~390 s the paper implies. Collection stopped at 1,119 because ~1,100 downloads in one sitting triggered YouTube's bot check — confirmed by ids already on disk failing too, i.e. throttling rather than link rot.
• Link rot. Of 2,000 sampled ids, 1,119 were retrieved. The true rot rate is below the implied 44% because later attempts were throttled, not dead.
• Clip length. Training with num_frames 16, sampling_rate 6 at 30 fps gives 3.20 s per clip (5.0 effective fps). The paper states 3.0 s / 5.33 fps, but its Table S1 also copies V-JEPA's sampling_rate=4, and those two are consistent only at a 21.33 fps source. 30/6 is 6.7% off the stated figure where 30 with stride 4 would be 29% off.

Usage

index.csv is a space-separated <relative_path> <label> list, the format facebookresearch/jepa's VideoDataset expects. The label column is unused by pretraining and is always 0.

from huggingface_hub import snapshot_download
root = snapshot_download("Ruian7P/HowTo1k", repo_type="dataset")
# rewrite index.csv to absolute paths before pointing a trainer at it

Licensing

The videos are YouTube content collected via the HowTo100M video list. HowTo100M itself is distributed as identifiers rather than media, and the authors' own video mirror is credential-gated. Redistribution here is for non-commercial research reproduction only; rights remain with the original uploaders.