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).
- Code: https://github.com/weathon/videojson_release
- LoRA adapter of the fine-tuned Qwen-3.5-Flash: https://huggingface.co/weathon/videojson_release_lora
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