主动交互原始视频

# Proactive Origin Data 20260529 This upload contains original/source video data and metadata manifests before colleague-side training preprocessing. ## Layout - `vqa/`: 7647 VQA QA rows from V-STaR and NExT-GQA. - `positive/`: 1515 proactive speech positive rows. - `negative/`: 3868 no-speech / silence negative rows. - `metadata/`: registry snapshots. Each subset contains: - `<subset>_origin_videos.tar`: source videos, archived under `<subset>/videos/`. - `<subset>_origin_videos.tar.sha256`: checksum. - `metadata/<subset>_origin_manifest.jsonl`: row-level metadata with `origin_video_archive_path`. - `metadata/<subset>_video_index.jsonl`: mapping from archive path to original server path. - `metadata/<subset>_summary.json`: counts and sizes. These are not the prepared frame datasets under `Data/raw/frames_*`.