docket-dataset-bench/social-norm-video-corpus

Social-Norm Video Corpus

A weakly supervised, multimodal corpus of social-norm events: short video intervals in which a concrete social behavior occurs, paired with evidence that the behavior is socially expected, disapproved of, corrected, praised, or otherwise norm-relevant. The corpus was mined from public video platforms in 2026. This release contains the full metadata, word-level transcripts and weak-label ledgers for the whole corpus, plus a small media sample that can be downloaded directly. The full media (hundreds of GB of clips) is available on request; see Access.

Status: v0.1, research preview. Labels are weak and audited only in part (see Label quality). Treat every label as a hypothesis, not ground truth. Media shards are being added progressively (October 2026); check media/*/index.jsonl for what is present.

Summary

Snapshot	corpus state of 2026-09-23 (label ledgers frozen 2026-08-11 to 2026-09-02)
Videos seen	351,083 enumerated; 318,620 with metadata; 269,509 with transcripts
Retained media	about 121,000 videos: 17,061 witnessed-reaction videos, 25,498 instructional, 33,527 commentary source videos, 30,475 negative-control videos
Weak-label rows	156,774 at the 2026-08-11 snapshot; 891,227 labeling-function votes over 171,508 gated items in the latest label run
Platforms	Dailymotion (about 85% of media, via its public Data API), Reddit (via the Arctic Shift API, v.redd.it media), Odysee/LBRY (Lighthouse search). YouTube was not used at scale.
Identifier	uid = "{platform}__{native_id}", unique across platforms; every table joins on it
Transcripts	WhisperX large-v3, word-level timestamps, language detected per video

The three evidence pillars

Pillar	What counts as evidence	Media	Weak-label rows (2026-08-11)
Witnessed	an on-scene response to a behavior: objection, correction, intervention, protective response, other contemporaneous reaction. Hits keep a window from about 12 s before to 2 s after the detected reaction phrase.	22,204 clips in 17,061 videos	32,955
Instructional	an explicit statement of a social rule paired with a demonstration (live action, role-play, animation, puppets, games, social experiments)	52,299 demo clips in 25,498 videos	64,888
Commentary	a narrator, participant or observer expresses a normative stance toward a concrete event. Visual supervision only when the event is present and localizable in the retained source video.	17,783 source videos retained of 33,527 catalogued	58,931
Negative controls	detector-clean 22 s windows from videos verified to contain no violation ("null_verified"), including same-source matched windows for 2,912 positive videos	29,175 videos	—

What is in this repository

README.md
metadata/
  videos.parquet                     one row per enumerated video (351K): uid, platform, url, title, channel, duration,
                                     status, query lineage, category, modality, pillar membership flags, has_transcript
  reactions.parquet                  61K detected reaction phrases: uid, clip_idx, phrase, tier, tag, start/end time, speaker, context
  queries.parquet                    52K search queries that drove collection, with per-query yield statistics
transcripts/
  transcripts-NNN.parquet            269K transcripts in shards of 10K: uid, language, n_segments, end_sec, text,
                                     segments_json (segments with word-level timestamps and confidence)
labels/
  witnessed_videos.parquet           per-video metadata for the witnessed pillar (category, provenance, scene labels)
  witnessed_reaction_text.parquet    the matched reaction utterance for each witnessed clip
  witnessed_norm_category_tags.parquet     21,809 clips: primary norm category, category counts, evidence tier
  witnessed_reaction_language_tags.parquet 21,809 clips: language and script of the reaction utterance
  witnessed_action_window_proposals.parquet 21,809 clips: proposed action window (about 10 s before the reaction),
                                     audio-impact snap, evidence tier (joint_high / norm_event_only / reaction_only)
  witnessed_action_clip_cut_manifest.parquet  20,905 cut action clips and their context clips, with sha256
  instructional_videos.parquet, instructional_demos.parquet   per-video and per-demonstration rows: the stated norm,
                                     polarity (follows / violates / explanation), quote anchors, time window, clip file
  instructional_demo_tier_manifest.parquet  4,679 demonstrations that passed the demo-tier review
  commentary_videos.parquet          33,527 commentary videos: category, agent (human / narrator), scene and provenance
  commentary_event_window_proposals.parquet 1,927 localized event windows in 943 commentary videos
  negative_videos.parquet, negative_windows.parquet   30,475 verified-null videos and their 22 s windows
  lf_gates.parquet                   171,508 items: eligibility gates for the label model
  lf_records.parquet                 891,227 labeling-function votes (lf_id, target, vote, confidence, evidence)
  ranked_norm_violation_candidates.parquet  the label model's ranked candidates (latest run)
  source_metadata.parquet            56,833 videos: platform-provided description and tags
  reddit_context.parquet             1,455 Reddit posts: title, score, kept comments (author names removed)
  *summary.json, dataset_views_rollup.json   run summaries for provenance
media/                               the clips, as tar shards of about 2 GB (gated; see Access)
  action_clips/*.tar + index.jsonl   20,905 cut action windows (~10 s before each witnessed reaction), one folder per uid inside each tar
  witnessed/*.tar + index.jsonl      36,256 witnessed hit clips (t-12 s .. t+2 s) with reaction text and per-video metadata
  instructional/*.tar + index.jsonl  96,056 demonstration clips with per-video metadata (norm statements, polarity, windows)
  negatives/*.tar + index.jsonl      123,546 verified-null 22 s windows with per-video metadata
                                     index.jsonl rows: {set, shard, uid, member, bytes}; members are "{uid}/{filename}"
annotation_pages/                    the browser-based review sheets used for human spot checks, self-contained
  spotcheck/index.html + clips/      "Norm Corpus Spotcheck v2": 132 clips across pillars with the questions reviewers answered
  paircheck/, paircheck_v2/          "Reaction Pair Baseline": 80 same-video pairs each; which 6 s clip precedes a reaction?
sample/
  index.jsonl                        what is in the sample and how big each item is
  witnessed_action_clips/{uid}/      40 short action clips (joint_high evidence) + reaction text, metadata, transcript
  instructional_demos/{uid}/         20 videos' demonstration clips + metadata + transcript
  negative_controls/{uid}/           15 videos' null windows + metadata + transcript
  commentary/{uid}/                  10 small commentary source videos + metadata + transcript

All nested JSON structures are stored as JSON strings in columns ending in _json or holding dict-valued fields; parse them with json.loads.

Label semantics and quality

The target, "a social-norm event occurred here", is latent. Labels come from tiered phrase and regex matching on transcripts, LLM scene labels, VLM frame labels, audio-event detection and source-context classification, combined in an append-only, audit-gated weak-supervision program:

• 28 weak-signal mechanisms are registered; 10 passed manual audits on frozen, source-disjoint cohorts for narrowly declared uses (retrieval, review ranking, strict-route exclusion, pillar rerouting); 18 failed to transfer and abstain.
• No rule is approved for automatic acceptance, and no rule deletes or rejects media.
• The latest label-model run (lf_records, lf_gates, ranked_norm_violation_candidates) is a Snorkel-style combination of those labeling functions; it has not been validated against a held-out human audit, which is why it ships as "proposals".

Known defects you should expect in first-pass labels:

• a relevant behavior is discussed but not shown;
• a scene is visible but does not match the assigned norm or polarity;
• instructional clips contain explanation, slides or B-roll instead of a connected demonstration;
• commentary text is valid but the event footage is absent or badly localized;
• witnessed clips show reactions to accidents, animals or spectacle without evidence of a social norm;
• the affected person, camera operator, authority or prank target is confused with a bystander;
• query lineage concentrates scripted series, compilations, news or police material.

Suggested tasks

1. Reaction prediction. From the clip before a reaction, predict reaction presence, responder role, strength and latency. Use negative_windows and the same-source matched negatives as controls.
2. Selection function. What gets reacted to, as a function of severity, head-count, norm category, content type and platform (witnessed_norm_category_tags, reactions, videos).
3. Enculturation channels. Does a model state a rule differently after seeing k clips from instruction, from observed sanction, or from commentary?
4. Counterfactuals. Matched pairs and minimal edits that flip a reaction predictor.

Report leakage probes (text-only, audio-only, OCR-only) beside every headline metric, and split by source (uid) and by channel; many videos share channels and scripted formats.

Access, licensing and ethics

• Media. The clips show identifiable people, often in unflattering moments. Access to this repository is gated: requests must state a research purpose and agree not to redistribute media or attempt to identify people. Commentary source videos and raw full-length videos are not included.
• Derived data. Our transcripts, labels and metadata are released for research use. The underlying videos remain the property of their creators and platforms; the release covers our derived annotations and short research excerpts only.
• Takedowns. Source URLs are kept for every item. If you are depicted in a clip or own a video and want it removed, open a discussion on this repository with the uid; items are removed and the snapshot is re-issued.
• Personal data. Platform usernames of commenters were removed from reddit_context. Channel names are retained because they identify public accounts, not private individuals.

Collection pipeline (for reproducibility)

queue of (platform, query) → enumerate candidates (Arctic Shift / Dailymotion API / Lighthouse) → deduplicate against state.db → duration pre-filter → download (yt-dlp) → transcribe (WhisperX large-v3) → detect reaction phrases (tiered phrase + regex matcher) → on a hit, cut the pre-reaction clip; otherwise purge the raw video → periodic query generation (exploit/explore per platform). Instructional and commentary pillars use the same pipeline with their own query families and LLM-based demonstration / stance labeling. Negative controls are sampled from videos the detectors mark clean and verified by a second pass.

Versioning and citation

This is v0.1 (2026-10). Later versions will add human-audited gold cohorts and a validated label model. If you use this corpus, please cite the dataset repository; a paper is in preparation and will be linked here. Questions and requests: open a discussion on this repository.