ckeith/meaningful-moments

Meaningful Moments (MM) — v1.0

Per-segment temporal-importance pseudo-labels for 536,181 videos across three canonical video action-recognition benchmarks, produced by querying a vision-language oracle (Qwen3-VL-32B-Instruct) once per video with a direct-scoring prompt. Each video is divided into fixed-duration segments; the oracle scores every segment's importance for recognizing the video's action, yielding ~4.58 million per-segment importance scores over 622 action classes.

substrate	train	validation	test	classes	Δt	prompt
Something-Something v2	168,913	24,777	27,157	174	0.5 s	direct_scoring_ssv2
Kinetics-400	239,789	19,877	38,671	400	1.0 s	direct_scoring_k400
Diving-48	15,027	1,970	—	48	0.5 s	direct_scoring_diving48

Source videos are not redistributed (see Licensing below). Every released label is keyed by the substrate's native video id.

What's in each instance

One row = one video. The conceptual instance of the datasheet is one (video, segment) pair — get that view with one explode:

from datasets import load_dataset
ds = load_dataset("<namespace>/meaningful-moments", "ssv2", split="validation")
df = ds.to_pandas().explode("segments")          # one row per (video, segment)

Top-level fields (identical across configs; template/placeholders extra on ssv2, class_id/raw_class_name extra on diving48, null where the source lacks them):

field	type	meaning
video_id	str	substrate-native video identifier
action_label	str	substrate class string (SSv2: filled caption; see template)
video_path	str	substrate-relative source path (SSv2/..., k400/..., diving48/...)
timestamp	str	ISO extraction time
summary	str	oracle's one-paragraph description of the video
label_counts	struct	per-video tallies of important/unimportant segments
segments	list<struct>	per-segment annotations (below)
mss_result	struct	extraction outcome (below)
elapsed_s	float	wall-clock for this video's annotation

Per segment:

field	type	meaning
index	int	0-based segment index
time_range	str	"t_start-t_end" in seconds
start_s, end_s	float	segment boundaries in seconds
label	str	important (kept in the minimal sufficient subset) or unimportant
weight	float	continuous importance in [0, 1] (= oracle's 0–100 score / 100)
phase	str	setup / contact / execution / continuation / result / disambiguation (null where the oracle response was unparseable)
reason	str	oracle's one-line rationale for the score

mss_result: kept_indices (the MSS as 0-based indices), total_oracle_calls (1 for direct scoring), precheck_passed (did the oracle confirm the action is present), precheck_vote_yes (vote fraction), precheck_responses[] (decision / confidence / evidence / rationale / YES-NO-SKIP logits).

Failure records are included

Videos where the oracle's precheck failed (action not confirmed: ~4.98% SSv2, ~7.80% K400, ~15.34% Diving-48) are released, not removed — they carry precheck_passed: false, empty kept_indices, and the oracle's response verbatim. Filter on mss_result.precheck_passed if you want clean labels only.

Repository layout

data/<config>/<split>-*.parquet      # this card's configs (raw_output omitted)
sidecars/<run>/sidecars-NNN.tar.gz   # canonical per-video JSON sidecars
sidecars/<run>/config.json           # full extraction provenance per run
supplement/cross_oracle/<oracle>/    # 4-oracle agreement pilot (600 videos
                                     #   each: Qwen3-VL-32B, Gemini 3.1 Pro,
                                     #   InternVL3-38B, GPT-5.5) — see its README
supplement/eval_sidecars/<run>/      # recognizer-eval records behind every
                                     #   thesis statistic (22 runs; added v1.1).
                                     #   With these + the code repo, every
                                     #   table/figure regenerates on a laptop —
                                     #   see eval_sidecar_runs.json for the map
                                     #   and sha256sums-eval.txt for integrity
manifests/<run>.csv                  # per-video index incl. shard membership
manifests/eval_pools/                # pinned stratified eval pools (seed=42)
prompts/                             # the four scoring prompts, verbatim
distributions.json                   # full-corpus per-substrate distributions
sample_sidecar.json                  # one canonical sidecar, for reference
sha256sums.txt                       # hash of every sidecar + released file
croissant.json                       # Croissant metadata

The JSON sidecars are the ground-truth format; parquet mirrors them field-for-field minus the verbatim raw_output string (recoverable from the tars). Verify any extracted sidecar against sha256sums.txt (sha256sum -c, paths are <run>/<video_id>.json).

Provenance

• Oracle: Qwen/Qwen3-VL-32B-Instruct, revision 0cfaf48183f594c314753d30a4c4974bc75f3ccb, greedy decoding (T=0).
• Prompts: SHA-256-pinned per run in sidecars/<run>/config.json; full text in prompts/.
• Pipeline: single-call direct scoring with a YES/NO/SKIP action-presence precheck; per-run hyperparameters (segment length Δt, mask operator) in each config.json.
• Cross-oracle supplement: the same pipeline run by three additional oracles over a shared 600-video pilot pool, for label-reliability analysis (supplement/cross_oracle/README.md).

Licensing

The MM labels (sidecars, parquet, manifests) are released under CC-BY-4.0. Source videos are governed by their own licenses and are not redistributed; obtain them upstream:

• Something-Something v2 — CC-BY-NC-4.0, from the Qualcomm AI Research distribution. (action_label/template strings on the ssv2 config derive from SSv2's annotations and carry that license's terms.)
• Kinetics-400 — CC-BY-4.0, from the DeepMind release.
• Diving-48 — research-only, from the UW-Madison distribution.

Versioning

• v1.0 — the label corpus (parquet + sidecars + cross-oracle pilot). Immutable.
• v1.1 — adds supplement/eval_sidecars/ (recognizer-eval records enabling GPU-free reproduction of every thesis statistic). No v1.0 file changed.

Future re-extractions under different oracles or prompt revisions will be released as sibling subsets under new semver tags — existing labels are never overwritten.

Code

The full annotation/evaluation/statistics pipeline: https://github.com/camkeith/meaningful-moments (MIT, tag v1.0) — see its REPRODUCING.md for the thesis-section → command map.

Citation

@misc{meaningfulmoments2026,
  title  = {Meaningful Moments: VLM-Derived Per-Segment Importance Labels for
            Video Action Recognition},
  author = {Keith, Cameron},
  year   = {2026},
  note   = {Dataset v1.0},
}

Contact: cameron.s.keith.26@dartmouth.edu — issues affecting downstream reproducibility answered within 7 days for the first 12 months post-release.