tianxia2/egolife-six-user-mcqa-textwrong131

EgoLife Six-User Multiple-Choice QA — text-only-filtered subset (131 questions, with frames)

A 131-question subset of tianxia2/egolife-six-user-mcqa-human-pass (169 human-verified five-option questions about 10-minute windows of EgoLife, where six people wear egocentric cameras at the same time; each question is asked by one person, the speaker). It ships the exact frames each question's generator received, plus the generator's evidence. A stricter 56-question subset is tianxia2/egolife-six-user-mcqa-hard56.

How the 131 were selected

Two models answered the 167 questions that have frames from the question text alone (same prompt, no frames; options shuffled with seed 0, greedy decoding): Qwen3-VL-8B-Instruct and Qwen2.5-VL-7B-Instruct. The 36 questions that both models answered correctly were removed; the 131 kept were answered wrong by at least one model (73 by both). selection.text_only_wrong_models lists which.

No filtering by human review. 29 of the questions were additionally re-reviewed by a human (questions that one model could answer from the speaker's own frames); the verdict is in selection.human_review_speaker_only (pass 12, fail 17, null = not re-reviewed). Drop the fail ones if you want only re-approved items.

Per-question baseline predictions (text only and speaker only) are in dataset.json → baseline_results.

Frames

For each question, frames lists, per person, the frames that were given to the question generator:

• 10-minute window sampled every 2 s → 300 frames per person (frame_XXXX_<seconds>s.jpg, seconds from window start).
• The speaker keeps all 300 frames.
• Each provider keeps only frames that are not redundant with the speaker's view: frames are CLIP-embedded and clustered per person; a provider cluster is dropped when it is very similar (cosine ≥ 0.82) to a speaker cluster within ±30 s; at least 40% of each provider's frames are kept. This depends on who the speaker is, so two questions on the same window with different speakers can use different provider frames. frame_counts gives the per-person totals.
• Files are the original 1408×1408 JPEGs, stored once per window and person: frames/<packet_id>/<agent_dir>/<file>. The generator itself saw them rescaled to 156×156 (area ≤ 24576), see meta.grpo_model_input_resolution.
• Frames carry a burned-in overlay with the day, clock time and person ID.

Full videos

videos/<packet_id>/<agent_dir>.mp4 are the original, un-pruned 10-minute videos of all six people for every window used here (414 files): 20 consecutive 30 s EgoLife clips concatenated, 640×640, 20 fps, ~600 s, with audio (the same files as in tianxia2/egolife-six-user-mcqa-human-pass). videos[*].path in the JSON files points to them. The frames above are the 2 s samples of the 1408×1408 source, after provider pruning; the videos contain every frame.

Files

File	Content
vqa.json	Question, options A–E, correct letter/answer, speaker, evidence_users / evidence_videos (two people cited by the generator, see the parent dataset), evidence_provider_users (the cited people other than the speaker: 2 for 63 questions, 1 for 67, none for 1), selection, frame_counts, frames.
vqa_shuffled.json	Same with options shuffled (random.Random(f"0:{id}")), as used in the evaluations; original_correct, option_order.
dataset.json	Full records: generator evidence (per-person claims and moments), original human labels, source clips, generator media metadata, baseline results.
video_manifest.json	One entry per video (review_video = path in videos/): its 20 source EgoLife clips (URLs) and the question ids in this subset that use it (used_by).
videos/	414 full 10-minute videos (6 per window).
frames/	89,255 JPEG frames.

Paths are relative to the repository root.