The Dataset Viewer has been disabled on this dataset.

VideoJSON artifacts

Artifacts of VideoJSON: A Minimal Harness Beats Purpose-Built Memory for Ultra-Long Video Understanding (paper).

Download into the root of the code repository:

hf download weathon/videojson_release --repo-type dataset --local-dir .
Path Content
data/egolife/<wearer>.json Minute index of the six EgoLife wearers: one JSON record per one-minute clip, captions by Qwen-3.8-Flash-Next
data/egolife/<wearer>_hourly.json Hourly index of the same recordings (+ Hourly Caption)
data/egolife/timelines/<wearer>.json Video files of every day with their start and end times, read by video_qa; the paths are those of our machine
data/mm_lifelong/{day,month}.json, {day,month}_hourly.json Minute and hourly index of MM-Lifelong Day and Month; MM-Lifelong Week uses data/egolife/A1_JAKE.json
data_cap35/ Minute index of A1_JAKE, Day and Month with captions by Qwen-3.5-Flash
mediastorm_livestream/month/train.json MediaStorm-Livestream: 400 open-ended questions in the MM-Lifelong format
mediastorm_livestream/minutes.json, hours.json Index of the livestream: one record per 30-second clip, and the hourly index
mediastorm_livestream/caption_schema.json, skill.txt, timeline.json, sandbox-image.json Caption schema, agent instructions, video timeline and sandbox image id that the harness reads with --external-month-dir
distill_data/train.jsonl, dev.jsonl, split.json, manifest.json Fine-tuning data: 352 training examples from 64 trajectories, the split of the questions by time, and the export manifest

The videos and the benchmark questions of EgoLife, MM-Lifelong, Ego-R1-Bench and EgoMemReason are not included; get them from their own releases. The livestream video is the public Bilibili replay https://www.bilibili.com/video/BV1rKzPBVEq1/.

Downloads last month
63