GMLRVigil/BenchCheck-MetaBenchmark

BenchCheck meta-benchmark (v5)

840 multiple-choice video questions from 76 public video benchmarks, one item per video, four capability groups x 210 (Perception, Temporal, Spatial / physical, Reasoning / knowledge). The set is the difficulty-first census of the BenchCheck screened pool: items that none of the cheap attackers of the pool screening (text-only, single frame, 32-frame 2B model, options-only, shuffled frames) could solve, ranked by the worst-case attacker percentile, with the marked answers verified against the frames and the reference model's share of correct answers capped.

Construction (five steps; details in the paper appendix)

1. Candidates: the screened pool; the 3,000 items with the smallest worst-case attacker percentile (one per video).
2. Capability labels: Claude Opus 5 assigns one of 18 fine categories from the question, options and gold answer; the categories map to the four groups.
3. Verification of the marked answer: Claude Opus 5 reads the frames (32, then 512 for unconfirmed items), the question, the options and the marked answer and returns supported / contradicted / not verifiable / definition dependent. Kept: supported; reversed clips (AoTBench RTime, TempCompass direction); not verifiable on videos >= 60 s. Dropped: contradicted; definition dependent; not verifiable on videos < 60 s. Ordering questions whose options are the events themselves are excluded (not scorable as single choice). 3b. Selection with a reference cap: per group the verified items in worst-case percentile order until 210, with at most 94 (45%) that the reference model (Claude Opus 5, VIDEO protocol) answers correctly.
4. Protocols: VIDEO (1 fps, <= 512 frames, short side 224) and BLIND (question + options, 4 permutations); see code/prompt.md.

Version v5: final composition.

Files

• items/meta_benchmark.jsonl: one line per item (fields: item_id, benchmark, capability_group, fine_category, question, options, answer, answer_idx, chance, video, video_id, duration_s, attack_pct, verification, source_version, license).
• items/summary.json: counts per group, benchmark and verification reason.
• videos/<video_id>.mp4: normalized videos (840 files; H.264, at most 1,024 stored frames per video).
• code/prompt.md: the exact prompts and the scoring rule.
• results/leaderboard.csv: current leaderboard (Overall = VIDEO, Blind, Gain, per-group accuracy; n = items scored).

Composition

Groups: {'perception': 210, 'temporal': 210, 'spatial_physical': 210, 'reasoning_knowledge': 210}. Largest sources: {'S-EMBER': 120, 'AoTBench': 87, 'ALLVB': 45, 'UrbanVideo-Bench': 34, 'TOC-Bench': 32, 'Video-MME-v2': 31, 'Camera Movement Understanding (BU_Google)': 31, 'VideoOdyssey': 30}. Verification reasons: {'reversed_clip': 85, 'supported': 500, 'not_verifiable_long': 255}. Median duration 52 s; 299 videos >= 300 s; mean chance 29.3%.

Licences

Every item keeps the licence of its source benchmark (field license; {'unknown': 183, 'CC-BY-4.0': 62, '': 42, 'custom': 62, 'CC-BY-NC-SA-4.0': 141, 'MIT': 103, 'CC-BY-NC-4.0': 161, 'Apache-2.0': 78, 'CC-BY-SA-4.0': 7, 'CC-BY-NC-ND-4.0': 1}). Items are redistributed for research evaluation only; check the source benchmark's terms before any other use.