VMemArena

Open-ended QA benchmark for long-video memory. 1000 questions over 561 videos (680 h of footage). Every question is free-form: there are no options to choose from.

Files

file	contents
vmemarena.json	561 video entries, each with its questions
videos.jsonl	per-video metadata (duration, fps, resolution, codec, size)
videos/	561 .mp4 files, named by video_id

Schema

// vmemarena.json — a list of video entries
{
  "video_id": "JmTtREhkJsA",
  "video_file": "videos/JmTtREhkJsA.mp4",
  "duration_s": 2139.97,
  "fps": 25.0,
  "has_audio": true,
  "questions": [
    {
      "qid": "CG-Bench:5477",
      "text": "How many people were left on the purple team when ...",
      "gt": "Four people.",
      "time_bin": "T1",              // retention span, T1 < T2 < T3 < T4
      "capability": "state_update",  // memory capability probed
      "scene_type": "multi_entity",  // visual scene character
      "source_dataset": "CG-Bench"
    }
  ]
}

The three axes

Retention span (time_bin) is balanced by construction — 250 questions each.

axis	values
time_bin	T1 250 · T2 250 · T3 250 · T4 250
capability	evidence_composition 254 · attribute_recall 253 · entity_tracking 251 · temporal_spatial_relation 186 · state_update 56
scene_type	multi_entity 380 · text_rich 275 · interaction_rich 218 · state_changing 127

capability and scene_type are reported rather than enforced: they reflect what the sources contain. state_update is supply-limited at 56 questions — treat per-class accuracy on it with the corresponding error bar, not as a headline number.

Notes on video_id

video_id is usually the source platform's id, with two exceptions that keep ids unique:

• <id>__<offset> — a clip, not the full video (17 files).
• <id>__<source> — the same platform id was collected independently by two source datasets and the two files genuinely differ, e.g. OFbyNU6UQQs__ScaleLong has no audio track while OFbyNU6UQQs__VideoOdyssey does. They count as two videos.

Provenance

Questions carry source_dataset. The 17 StreamArena-HR questions passed a separate audit from the other 983; the two review pipelines were never calibrated against each other, so filter on source_dataset if that matters for your analysis.

Videos belong to their original publishers and are redistributed here for internal evaluation only.