Meaningful Moments — code release

Annotation, evaluation, statistics, and plotting pipeline behind Meaningful Moments (MM): ~4.58M per-segment temporal-importance pseudo-labels over 536,181 videos across SSv2, Kinetics-400, and Diving-48, produced by a Qwen3-VL-32B oracle with single-call direct scoring.

• Dataset: https://huggingface.co/datasets/ckeith/meaningful-moments (CC-BY-4.0, tag v1.0)
• Code license: MIT (this repository)
• Reproduction guide: REPRODUCING.md — every thesis table/figure mapped to the command that regenerates it

Install

python3 -m venv venv && source venv/bin/activate   # Python 3.12 used in production
pip install -r oracle/requirements.txt             # curated direct deps
# or, for the exact production environment:
pip install -r requirements-freeze.txt

Set MM_ROOT to this repository's root when running from elsewhere (defaults resolve relative to each script's location).

Quickstart (Tier 1 — no GPUs, released data only)

from datasets import load_dataset
ds = load_dataset("ckeith/meaningful-moments", "ssv2", split="validation")
print(ds[0]["segments"])     # per-segment weight / phase / reason

Regenerate a thesis analysis from the released eval-sidecar supplement (see REPRODUCING.md §Tier-1 for fetching it):

python scripts/mode_prompt_agreement.py --out /tmp/mm_check
# reproduces the mode/prompt-ablation agreement tables exactly

Annotate your own video (Tier 2 — needs a GPU)

python oracle/scripts/mss_extract.py \
    --video your_video.mp4 --label "your action label" \
    --output out.json --model qwen3-vl-32b --direct-scoring

Layout

oracle/scripts/mss/        oracle, segmentation, masking, prompts (the annotator)
oracle/scripts/            evaluation CLIs (recognizer eval, paired stats, I/D curves,
                           shuffle damage, temporal coverage, cross-oracle, heads)
scripts/                   analysis wrappers (exp5 CIs, mode/prompt agreement,
                           galleries, eval-pool samplers)
scripts/hf_release/        the pipeline that built the HF dataset release
scripts/distributions/     corpus statistics extractor
paper_tables_*/            figure plotters
data/csvs/                 pinned evaluation pools (seed=42)

Citation

See CITATION.cff.