What we kept, what we rejected, and how to check
A dataset is a set of decisions. This page documents ours.
Cheap filters first, expensive ones last
Most candidates are rejected from metadata, thumbnails and embeddings alone, before a single full video is downloaded. Pass rates below are from a recent production run.
- S0 79% Metadata gates pass rate on duration, language and live-status checks — the cheapest filters run first
- S1 82% Frame & embedding probes storyboard frames + viewpoint/domain embedding classification, still with no full download
- S2 33% Shortlist for full download only top-scoring candidates are ever fetched at full resolution — 732 of the 3,479 videos discovered
- S3 ×5 Lossless cutting each downloaded video yields multiple candidate procedure clips
- S4 ≈13% Final curation of cut clips survive balancing by domain, viewpoint, resolution and leakage screening
We collect only from the open public web, and every record documents its origin. When your compliance team wants specifics (sources, rights posture, jurisdictions), we answer directly, as part of scoping.
“Looks fine” is not a pass
Every dataset line defines explicit, machine-checkable gates. Example — our talking-head line, where the bar is: the visible mouth must be producing the audible sound.
- Dubbed audio over original footage — lips don't produce the sound
- Narrator voice over silent B-roll faces
- AI-generated avatars presented as real speakers
- Codec-level audio/video desync
Each record stores its per-gate verdict, so you can audit not just what was delivered but why it was kept. In acceptance testing, deliberately poisoned probes were rejected 4/4: dubbing and language mismatch by the automated sync and language gates, and a synthetic avatar by the curated blocklist inside the authenticity gate. We document which defense caught what.
Model output, measured like model output
Automatic annotation is only trustworthy with a measured error bar. We maintain human-labeled golden sets and report agreement on every annotation run.
Clean evals require clean provenance
Every source video ID is screened against the public corpora your model may already have trained on. For video: COIN, CrossTask, YouCook2, HowTo100M, HTM-AA, VidChapters and others — 2.02M+ IDs, which removed 2.4% of candidates from the reference batch. For speech: AVSpeech, VoxCeleb1/2, VoxBlink+2, VoxLingua107, CelebV-HQ, CelebV-Text, TalkingHead-1KH and HDTF — 3.21M IDs, with zero overlap found in the current delivery. That is 15 public corpora and 5.2M+ IDs across the two lines. If you build evals, this is the difference between a benchmark and a leak.
Chain of custody, per record
- source URL
- exact license
- fetch timestamp
- transform hash
- dataset card / datasheet
- QA report with golden-set metrics
- SHA-256 manifest, verified at acceptance
Deliveries are individually fingerprinted, and due-diligence questionnaires are answered on request. License filtering to your constraints — for example, permissively-licensed subsets only — is part of scoping.
The questions your counsel will ask
Rights review is a stage of every serious data purchase. We build for it up front so it doesn't stall your timeline later.
- Origin on every record — source, license, fetch time, transform hash — plus a signed manifest on every delivery.
- License filtering to your constraints (for example, permissively-licensed subsets only) is part of scoping.
- People-centric lines are screened during sourcing: content featuring minors or synthetic media is removed, and identifiable-person data ships under terms agreed per engagement.
- Biometric and consent posture for facial-landmark data is covered in the due-diligence questionnaire, mapped to the regional regimes that apply to you.
- Due-diligence questionnaires answered on request; datasheets document known limitations, including the unflattering ones.