Dataset: v1.1.0
English | 简体中文
Visual-only timestamped QA and Proactive Response annotations derived from StreamingBench and OVO-Bench, with reviewed temporal corrections.
| Population | Size |
|---|---|
| QA | 833 records |
| Proactive | 415 records / 1,338 windows |
| Proactive standard | 407 records / 1,270 windows |
| Proactive sampling stress | 8 records / 68 windows |
| Tiny setup subset | 9 QA + 9 Proactive records |
--subset full includes sampling stress. Tiny checks setup, not model ranking.
Data and software versions are independent.
Files and time coordinates
See the release file and field reference for every file and its fields.
qa.jsonl and proactive.jsonl contain canonical records. ID files define
subsets. manifest.json contains hashes and source fingerprints; source paths
are provenance identifiers, not required local files. Audit sidecars document
annotation changes and exclusions.
Timestamps refer to the original video. QA records contain question, options,
answer, question time and evidence anchor. Proactive records contain instruction
time, event/state annotations and answers. Legacy deadline_s is provenance,
not a cutoff. See the protocol.
Limitations and access
525 records were human-reviewed; 723 passed model-assisted screening without individual human review. Do not describe all records as human-validated. Dense windows may be unobservable at 1 FPS; report coverage and sampling-stress results separately. Public GT limits claims about unseen-data generalization.
No videos or weights are included. See data setup. The StreamingBench authors have permitted redistribution of the modified annotation files in this repository; StreamingBench-derived TRACE additions are offered under CC BY-NC-SA 4.0 within the rights held by TRACE contributors. This does not license the original videos or upstream source fields. OVO-Bench annotation terms remain pending; see data terms.