BootsofLagrangian/V-JEPA-Reproduce-Recipe

V-JEPA Reproduce Recipe

Private, metadata-only planning surface for reconstructing a large-scale V-JEPA-style observation-pretraining mixture. This repository does not redistribute source videos and does not grant rights beyond each upstream dataset's terms.

Current release

• Manifest ID: YTWS-CANONICAL-MANIFEST-V1
• Run ID: 823483f781094a668cf6682828723d16
• Schema version: 1
• Source memberships: 485,265,004
• Exact interval samples: 469,907,459
• Canonical assets: 82,701,705
• Parquet shards: 3,292
• Parquet-only payload: 41,557,148,598 bytes (38.70 GiB)

The manifest combines the 11 currently materialized metadata sources: COIN, DROID, EPIC-KITCHENS-100, HD-VILA-100M, HowTo100M ID recovery (unattested ID mirror), InternVid, Kinetics-710, OpenVid-1M, Panda-70M, YouNiverse (structural HF mirror), and YT-Temporal-180M (unverified mirror). Provenance quality is recorded per source. Dataset membership is not a statement that every media URL remains reachable.

Layout

manifest.json
samples/dedup_bucket=00..ff/*.parquet
assets/dedup_bucket=00..ff/*.parquet
memberships/dataset_id=<name>/dedup_bucket=00..ff/*.parquet
dataset_roles.parquet
dataset_stats.parquet
source_coverage.parquet

• samples: exact (platform, video_id, millisecond interval) identities.
• assets: source video or trajectory identities without interval duplication.
• memberships: lossless lineage back to each source metadata row.
• dataset_roles: the overlapping broad_in_the_wild, instructional, curated_temporal_action, ego_and_synchronized_observation, and hosted_quality_video views.
• dataset_stats: per-source row, dedup, reservation, and resolvability counts.
• source_coverage: included and excluded catalog sources with provenance or exclusion reasons.

Captions, action labels, and other source-specific payload tables are not part of this repository. Membership rows retain dataset_id, source_row_id, and metadata_revision as lineage keys for consumers that separately acquired the corresponding upstream payload. The HowTo100M recovery contains IDs only; the canonical captions/tasks archive was unavailable and cannot be reconstructed from this release.

Load with Hugging Face Datasets

Authentication is required because the repository is private; export a token with read access as HF_TOKEN. The repository exposes six named configurations. samples is the default, so both the bare call and an explicit "samples" select the deduplicated interval table. Use "assets" and "memberships" for the other large tables; "dataset_roles", "dataset_stats", and "source_coverage" expose the three compact indexes.

The simplest streaming load is:

from datasets import load_dataset

repo_id = "BootsofLagrangian/V-JEPA-Reproduce-Recipe"
samples = load_dataset(repo_id, "samples", split="train", streaming=True)

web_train = samples.filter(
    lambda row: (
        "broad_in_the_wild" in row["roles"]
        and not row["is_reserved"]
        and row["directly_resolvable"]
    )
)

For source lineage, load the memberships configuration and filter by dataset_id. To minimize shard discovery and assign the 256 hash buckets directly across workers, address only the required source and bucket paths:

dataset_id = "howto100m"
bucket = "00"
memberships = load_dataset(
    "parquet",
    data_files={
        "train": (
            f"hf://datasets/{repo_id}/memberships/"
            f"dataset_id={dataset_id}/dedup_bucket={bucket}/*.parquet"
        )
    },
    split="train",
    streaming=True,
)

The same direct form applies to samples/dedup_bucket=<bucket>/*.parquet and assets/dedup_bucket=<bucket>/*.parquet.

revision= is optional for ordinary use because main is the current release. For reproducible training, pass the immutable data commit shown below as revision="4b9c79918adcb426fa7d0b27056fcde9366357af" to load_dataset, or add @<revision> to an hf:// URL. The manifest run ID remains inside manifest.json as provenance; it is not part of the Hub path.

The repository-level V-JEPA-scale pretraining guide records the media-materialization boundary, capacity estimates, compute profiles, and launch gates. The guide requires access to the private GitHub repository; this dataset card remains the data-release contract.

Exact-dedup boundary

This release removes exact repeated memberships at identical platform, video identity, and millisecond interval keys. It does not merge merely overlapping clips and does not claim perceptual or semantic deduplication. Whole-asset reservation propagation prevents a held-out asset from leaking through another dataset membership.

Publication verification

Data commit 4b9c79918adcb426fa7d0b27056fcde9366357af is the immutable revision for this release. The published tree was checked as 3,293 data objects totaling 41,557,605,070 bytes: 3,292 Parquet shards plus manifest.json. Every local object size and SHA-256 matched its remote Hub object. Pinned streaming loads each returned a row from the samples, assets, and memberships configurations; the bare call also resolved to the default samples config. The earlier data commit e3b4f02a422f760995406a0b8e1221e3882ec7d7 is superseded: its runs/<run_id> paths remain available in this Hub repository's Git history but no longer resolve on main.

Media smoke status

One fixed random membership from each included source was probed locally. Six of eleven produced MP4 files that passed ffprobe and one-frame CPU decode. Four YouTube-backed rows reached format extraction but failed at the bounded ffmpeg section-request boundary with HTTP 403; the selected EPIC remote seek timed out. This tiny smoke is an integration check, not an estimate of live URL fraction or dataset quality.

Planned follow-up

• Extend the pinned streaming-load check from the exercised representative samples, assets, and memberships shards to all 256 sample and asset buckets and every membership partition.
• Add DINOv2-L embeddings and a separately versioned perceptual/retrieval dedup layer without rewriting the exact-identity tables.
• Define replacement sampling weights only after coverage, duration, role, and availability measurements; do not treat source row counts as mixture weights.
• Retry the same failed YouTube smoke IDs with an official yt-dlp EJS runtime and a bounded native-download-then-local-trim fallback.
• Add a range-aware EPIC media adapter and repeat the same selected ID.
• Run a separately registered, bounded, stratified random-asset reachability measurement for every included source. Preserve the selected asset as the inferential unit, prohibit candidate replacement, and report per-source success fractions with the egress, timestamp, adapter, and uncertainty.
• Recover or explicitly retain exclusions for unavailable, gated, or legally withheld sources, including YT-Temporal-1B, LVD-142M, SSv2, Ego4D, Ego-Exo4D, and WebVid.
• Decide separately whether source-specific captions and annotations should be published as joinable private tables. They are not required for the current observation-pretraining manifest.
• Keep post-training/evaluation manifests separate from this pretraining release; they are intentionally not part of this upload.
• Add video-storage or external-stream execution only as a versioned successor after media licensing, availability, codec, and GPU decode policies are fixed.

License and access

The repository must remain private while provenance and redistribution terms are audited. Each source retains its own license and access restrictions. A working public URL or community mirror is not, by itself, redistribution permission for metadata or media.