ProReady-QA

A benchmark for when a video model should answer, not just what it answers.

Every question is posed at a timestamp question_timestamp and its answer is not yet knowable at that moment. The supporting evidence appears later, inside an annotated window [evidence_start, evidence_end]. A model that answers before the window is speculating; one that answers long after it has stopped being useful in a streaming setting.

Introduced in StreamReady: Learning What to Answer and When in Long Streaming Videos (CVPR 2026), arXiv:2603.08620.

Evaluation-only. There is no train split: the benchmark measures answer timing, and training on it would defeat the purpose.

Built on VStream-QA (RVS split), and shipped in the same flat per-source layout so it drops onto an existing VStream-QA frame extraction without renaming anything.

At a glance

	Ego4D	MovieNet	total
videos	10	22	32
QA pairs	1,828	2,679	4,507
dialogues	1,802	2,614	4,416
multi-turn dialogues	21	52	73
τ (median evidence duration)	29	89	—
median lead time	151	181	—
minimum lead time	30	90	—

task	Ego4D	MovieNet		difficulty	Ego4D	MovieNet
SSR Sequential Steps Recognition	545	797		easy	527	751
GSD Goal-State Detection	364	601		medium	940	1,338
REC Repetitive Event Count	343	500		hard	361	590
CRR Clues Reveal Responding	317	480
CTD Causal Trigger Detection	259	301

• SSR — what step comes next in a process already underway
• REC — how many times a recurring event happens (answer is an integer)
• CRR — answerable once a specific later clue appears
• CTD — what will cause an upcoming event
• GSD — whether an activity reaches its goal state

Files

proready_qa_ego4d.json      0.91 MB   1,828 QA — flat list of records
proready_qa_movienet.json   1.38 MB   2,679 QA — flat list of records
frames_manifest.json                  per-video frame counts + digests
clips/ego.json                        the 10 Ego4D clips to cut (seconds)
clips/movie.json                      the 22 MovieNet clips to cut (shots)

Video frames are not included. Ego4D requires each user to accept its own licence agreement, so redistributing pixels is not permitted.

Schema

A flat list of records:

{
  "video_id": "000012",
  "id": "000012_000000",
  "question": "What does the person add next to the tortilla?",
  "answer": "Grated cheese",
  "answer_type": "SSR",
  "video_name": "9198b9a4-8d8f-4ba6-9924-4c86982d890a",
  "source": "ego4d",
  "question_timestamp": 630.0,
  "evidence_start": 811.0,
  "evidence_end": 840.0,
  "difficulty": "easy",
  "duration": 3600.0,
  "dialogue_id": "000012_d0000",
  "turn_index": 0,
  "multi_turn": false
}

field	meaning
question_timestamp	when the question is asked
evidence_start / evidence_end	the verified window in which the answer becomes visible
dialogue_id / turn_index	multi-turn grouping; groupby("dialogue_id") recovers dialogues
references_qa_id	present on follow-up turns only, pointing at the earlier turn they refer to

id is globally unique (it embeds video_id), so it can be used as a single join key.

evidence_start > question_timestamp holds for every record — that is the proactive requirement, and it is enforced at export.

import json, pandas as pd
df = pd.DataFrame(json.load(open("proready_qa_ego4d.json")))
df.groupby("dialogue_id").size().value_counts()   # dialogue lengths

Getting the frames

1. Get the source data. Follow the official instructions for Ego4D (licence agreement required) and MovieNet (keyframes).

2. Cut the clips. clips/ego.json gives original_video, start_time, end_time in seconds; clips/movie.json gives start_shot, end_shot. Save each as <video_id>.

3. Extract frames at 1 FPS into this layout, 1-indexed:

frames/<video_id>/frame000001.jpg      # bare 000001.jpg also works

4. Verify against frames_manifest.json, which records frame_count and a frame_digest over sorted (filename, size) pairs. A count mismatch is the usual symptom of extracting at the wrong rate.

The 1 FPS rate is load-bearing. Timestamps are in the same units as frame indices. Extract at a different rate and every annotation lands on the wrong frame unless you rescale.

License

CC BY-NC 4.0, inherited from VStream-QA. Non-commercial research use.

Using this benchmark also means complying with the terms of the underlying sources, which this license does not waive:

• Ego4D — requires accepting the Ego4D license agreement
• MovieNet — see MovieNet terms
• VStream-QA — CC BY-NC 4.0

Citation

@inproceedings{azad2026streamready,
  title     = {StreamReady: Learning What to Answer and When in Long Streaming Videos},
  author    = {Azad, Shehreen and Vineet, Vibhav and Rawat, Yogesh Singh},
  booktitle = {CVPR},
  year      = {2026}
}