happy8825/videosearch-r1-demo-examples

VideoSearch-R1 demo examples

23 hand-picked qualitative examples from VideoSearch-R1: Iterative Video Retrieval and Reasoning via Soft Query Refinement (ECCV 2026, arXiv:2607.00446). Each example has the video files the model saw, the text query, the ground-truth moment, and the model's full multi-turn log (reasoning, search, verification, grounding, final answer).

Model outputs are verbatim eval logs of the released checkpoints VideoSearchR1/didemo-stage2 and VideoSearchR1/charades-stage2. Nothing was re-run.

	DiDeMo	Charades
main (turn-1 retrieval wrong, fixed by refinement)	8	8
single (matched at turn 1)	2	2
optional (edge cases)	1	2

Files

examples.json            23 examples, nested: {meta, examples[]}  ← canonical source
data/examples.jsonl      same 23 examples, one row each, flattened for the Dataset Viewer
data/turns.jsonl         43 rows, one per (example, turn) — handy for drawing timelines
videos/didemo/*.mp4      20 DiDeMo videos  (Flickr / YFCC100M)
videos/charades/*.mp4    20 Charades videos (Charades_v1_480)
thumbs/<id>.jpg          one frame from the middle of the GT moment, 480 px wide
preview_sheets/*.jpg     contact sheets used to pick the examples (green box = GT moment, blue bar = prediction)
index.html               static browser for the 23 picks (open locally next to the other files)

Videos that were only the wrong rank-1 result at turn 1 are included too, so a demo can show what the model looked at before refining.

Load

from datasets import load_dataset
ex    = load_dataset("happy8825/videosearch-r1-demo-examples", "examples", split="train")
turns = load_dataset("happy8825/videosearch-r1-demo-examples", "turns",    split="train")

from huggingface_hub import snapshot_download
root = snapshot_download("happy8825/videosearch-r1-demo-examples", repo_type="dataset")  # videos + json

The task

Video Corpus Moment Retrieval. Given a text query:

1. Search: the search engine (Qwen3-VL-Embedding-2B) returns the rank-1 video from the whole test corpus (DiDeMo 1,002 videos, Charades 1,334 videos).
2. Verify: the policy (Qwen3-VL-4B) watches that video, writes <think>…</think> and answers matched / not_matched.
3. Soft Query Refinement: on not_matched it emits <REFINE>; 8 soft query tokens are generated in latent space, appended to the query embedding, and the search is re-run. Max 3 turns on DiDeMo, 2 on Charades.
4. Ground: on matched it predicts <start> / <end> of the moment in seconds.

Schema: data/examples.jsonl (one row per example)

column	meaning
id, dataset, kind, qid	id = <dataset>-<index>; kind ∈ main / single / optional
question	the text query
gt_video, gt_file, gt_start_sec, gt_end_sec	ground-truth video and moment (seconds)
initial_retrieval_rank_of_gt	rank of the GT video with the plain query (turn-1 search)
num_turns, max_turns, stop_reason	how many turns ran and why it stopped
final_video, final_is_ground_truth, final_start_sec, final_end_sec, final_iou_with_gt	the pipeline's final answer
videos[]	every video the model saw: video, file, duration_sec, is_ground_truth
turns[]	the full per-turn log (same objects as in turns.jsonl, nested)
thumbnail	path to thumbs/<id>.jpg

Schema: data/turns.jsonl (one row per turn)

column	meaning
id, dataset, kind, question, turn	which example and which turn (1-based)
search_query	turn 1: the text query. turn ≥ 2: the refined query is latent (8 soft tokens) and has no text form, so this is a description
query_refined_from_previous_turn	true for turn ≥ 2
retrieved_rank, retrieved_video, retrieved_file, retrieved_is_ground_truth	the rank-1 video the search engine returned this turn
retrieved_score	always null: similarity scores and the rest of the top-k are not in the log
ground_truth_rank_after_this_search	where the GT video ranked after this turn's search (shows the refinement effect, e.g. 91 → 1)
reasoning	contents of <think>
verification	matched / not_matched
grounding_start_sec, grounding_end_sec	predicted moment on a match, else null
iou_with_gt	IoU of that prediction with the GT moment
emits_refine_token	true when the turn ended with <REFINE>
raw_output	the model's full text, for display
gt_video, gt_start_sec, gt_end_sec	copied from the example for convenience

Timeline note. The search engine retrieves whole videos, so there are no retrieved segments per turn. The time spans to draw are gt_start_sec/gt_end_sec and each turn's grounding_start_sec/grounding_end_sec. All values are seconds; no fps conversion is needed.

Picks at a glance

id	query (short)	turn-1 result → why rejected	GT rank 1→2	final IoU
didemo-3578	man in red shirt points at camera on a Segway	park crowd, pink shirt, no Segway	8 → 1	1.00
didemo-620	tuba players lift horns together	saxophone players	2 → 1	1.00
didemo-1721	hand taps a ferret three times	a cat being petted	3 → 1	1.00
didemo-2373	yellow fish swims left, bubbles	koi pond, no yellow fish	3 → 1	1.00
didemo-2784	blonde girl kisses baby in red shirt	baby in white, no red shirt	5 → 1	1.00
didemo-246	trolley arrives from the right	train arrives from the left	2 → 1	1.00
didemo-1485	man in green shirt celebrates at baseball game	man in red shirt celebrating	2 → 1	1.00
didemo-2516	student with pink umbrella cheering	marching-band crowd	5 → 1	1.00
didemo-21	chinchilla jumps off perch	— (matched at turn 1)	1	1.00
didemo-3	tennis players shake hands at the net	— (matched at turn 1)	1	1.00
charades-3322	lots of cooking on a white electric stove	other kitchen, not "a lot of cooking"	91 → 1	0.98
charades-2290	white t-shirt, laptop in a chair	pink shirt at a desk	47 → 1	0.85
charades-1736	reading a book, blue wall behind	red curtains behind	23 → 1	0.88
charades-1379	yellow hoodie opens refrigerator	black clothes at refrigerator	12 → 1	0.85
charades-1669	black cap + gray shirt takes a bite	white cap + blue shirt	4 → 1	0.94
charades-2831	closes door holding a yellow bag	green bag, no yellow bag	2 → 1	1.00
charades-2478	stretches on green couch, woman in room	no woman present	2 → 1	0.88
charades-19	white shirt + jeans at a table with papers	blue top at a round table	6 → 1	0.93
charades-1237	pours water in front of a monitor	— (matched at turn 1)	1	0.98
charades-341	pink hoodie + black cap eats with a fork	— (matched at turn 1)	1	0.96

Optional extras: didemo-1522 (the one 3-turn success: GT retrieved at turn 2 but rejected, accepted at turn 3), charades-2703, charades-909 (shares both videos with charades-2290; use one or the other).

Notes

• The trailing <REFINE> on matched turns is an artifact of the output format; the pipeline stops on matched and ignores it.
• DiDeMo moments are 5-second chunks and videos were truncated to 30 s during annotation, so GT spans are multiples of 5 s even when the file is longer.
• Sources: DiDeMo-FIG and Charades-FIG test splits, eval/external_verified_test_temporal_grounding.check.jsonl in each checkpoint repo.

Citation

@inproceedings{videosearchr1_2026,
  title     = {VideoSearch-R1: Iterative Video Retrieval and Reasoning via Soft Query Refinement},
  booktitle = {ECCV},
  year      = {2026},
  eprint    = {2607.00446},
  archivePrefix = {arXiv}
}