DIVE-Bench

Two separately evaluated scenarios, named after sections 4.2 and 4.3 of the revised manuscript distributed with the public code:

Paper scenario	Hugging Face configuration	Test examples	Identity
Educational High-FPS Videos	educational_high_fps	634 QA rows over 317 videos	Unique qid
High-Motion High-FPS Videos	high_motion_high_fps	3,243 clips	(video_path, qid); numeric qids alone repeat

The configurations are subsets, not train/test splits of a single task. Each has its own test split, references and metrics. Do not concatenate their scores or average their unrelated primary metrics. The earlier arXiv paper describes the educational version; it should not be presented as containing the later two-scenario release.

Load the two tasks

from datasets import load_dataset

educational = load_dataset(
    "haichaozhang/DIVE-Bench", "educational_high_fps", split="test", token=True
)
high_motion = load_dataset(
    "haichaozhang/DIVE-Bench", "high_motion_high_fps", split="test", token=True
)
assert len(educational) == 634
assert len(high_motion) == 3243

# Historical preview profile, not a third HF subtask:
high_motion_preview1000 = high_motion.select(range(1000))

Access controls and source terms still apply. Pin the repository revision when reporting results. Video assets are not embedded in these annotation files and must be obtained with the appropriate source access. Local annotation loading does not prove that a fresh public caller can download the source videos.

Frozen annotations and compatibility

Annotation bytes, columns, qids and row order are preserved. In particular, LPM_slides.parquet from the old educational repository is an identical duplicate of LPM_videos.parquet; it must not be loaded as additional examples.

File	SHA-256
data/educational_high_fps/LPM_videos.parquet	9ab09ea35a66ca86fc7fdce4e539171ea5eef38f8162874f74dd1b2809dafb66
data/high_motion_high_fps/Egodex_traj.parquet	518e2896749b4d6e957d7e9fb0ae16f75c28954e50ef84303889070253cf8ecd

The high-motion file preserves the source Hub's original Parquet bytes. The historical local release also has a 39f9da7a... Parquet serialization. Their ordered task content is identical: SHA-256 90ee915016105f6a709f391e8a03a6d0e99bc5c908f945cdf7b80d0cb289e789 over [video_path, qid, question, answer, int(frame_count)] for every row, encoded as compact UTF-8 JSON (ensure_ascii=False, separators (',', ':')).

Existing repositories and their historical versions remain separate compatibility sources; they are not deleted or silently rewritten by this organization:

• Educational source, revision 5cc61a045c8e5e95d1d9c87e22ccd0f699575aea.
• High-motion source. Revision d44407f607fdf020c59b816884f06ed6d453cf26 was verified through the owner account. The source repository is private; this is not a public-access claim.

Video paths and frame policy

Preserve educational DenseVideo-LPM/videos/<video-id>.mp4 paths and high-motion egodex/<action>/<numeric-id>.mp4 paths under an explicit local data root. Never flatten high-motion filenames: numeric ids repeat across 111 action directories.

High-motion answer is a JSON list of grid labels and answer_traj is a JSON coordinate sequence. Both span all frame_count source frames. When evaluating with K sampled frames, sample references and input frames at the same endpoint-inclusive integer linspace(0, F-1, min(K,F)) indices. A nominal nframe=8 setting is not evidence that a wrapper actually used those frames. Report decoded frame counts, positions and any clipping/truncation policy.

The historical leaderboard preview uses the first 1,000 annotation rows. Retain this order exactly. Full 3,243-row evaluations must be labeled separately. The High-FPS scenario names do not imply that a model processed every source frame; the default reference operating point samples eight frames.

Evaluation and release status

Educational objective metrics: per-example CER, WER, Token-F1 and Exact Match. High-motion objective metrics: grid accuracy, normalized grid-center ADE/FDE, transition accuracy and canonical-label Token-F1. Open MOS is a separate pinned judge protocol, not an automatic field in these data files.

Use the public code's reproduction contract and score audit. Framework submissions are draft integrations, not upstream acceptance: VLMEvalKit #1686 and lmms-eval #1521.

See source-specific terms before using or redistributing data. Organizing configurations does not grant additional rights or prove a fresh GPU/judge reproduction of historical scores.