File size: 4,501 Bytes
05a25c0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
---
language:
- en
pretty_name: "VDR: Video Deep Research Benchmark"
tags:
- video
- video-question-answering
- multimodal
- deep-research
- multi-hop-reasoning
- agents
task_categories:
- visual-question-answering
- video-text-to-text
size_categories:
- n<1K
configs:
- config_name: default
  data_files:
  - split: test
    path: data/vdr_questions.parquet
---

# VDR: Video Deep Research Benchmark

**VDR (Video Deep Research)** is an evaluation benchmark for *video-grounded deep
research*. Each item pairs a video with a question that is deliberately constructed
so that **neither the video alone nor the open web alone is sufficient** to answer it:
a system must extract concrete observations from the video, use them as search anchors
to retrieve external information, and combine both to reach the answer.

- **Videos:** 49
- **Questions:** 100
- **Split:** `test` (evaluation only — the released items are *sanitized* and contain no ground-truth answers)
- **Language:** English

## Dataset structure

The questions table is published as Parquet for easy browsing and loading:

```
data/vdr_questions.parquet   # 100 rows, one per question
videos/<video_id>.mp4        # 49 source videos (referenced by the table, not duplicated into Parquet)
questions/<video_id>/<qid>.json
video_question_map.json      # index of questions grouped by video
```

Videos are **not embedded** in the Parquet. Each row references its video by a
repo-relative path (`video`) and a directly resolvable link (`video_url`), following
the common pattern for larger video datasets where media is stored as files rather
than inlined bytes.

### Data fields

| field | type | description |
|-------|------|-------------|
| `sample_id` | string | Unique item identifier to report in predictions (e.g. `2PXEEdsJbnQ_q1`). |
| `video_id` | string | Source video identifier (e.g. `2PXEEdsJbnQ`). |
| `question` | string | The question to answer (source field: `task_question`). |
| `video` | string | Repo-relative path to the MP4, e.g. `videos/<video_id>.mp4`. |
| `video_url` | string | Directly resolvable URL to the MP4 on the Hub. |
| `audio_required` | bool | Whether answering requires the audio track. |
| `image_search_required` | bool | Whether answering requires image/visual web search. |

### Statistics

- 49 videos, 100 questions (1–3 questions per video: 14 videos with 1, 19 with 2, 16 with 3).
- Total video size ≈ 4.77 GB (per-video min / median / max ≈ 2.8 / 25.1 / 1534.6 MB).
- `audio_required`: 7 questions. `image_search_required`: 6 questions.

## Usage

```python
from datasets import load_dataset

ds = load_dataset("hh721/vdr_temp", split="test")
print(ds[0]["sample_id"], ds[0]["question"])
```

Fetching a video for an item:

```python
import urllib.request

item = ds[0]
urllib.request.urlretrieve(item["video_url"], f'{item["video_id"]}.mp4')
# or use the repo-relative path with huggingface_hub.hf_hub_download(..., item["video"])
```

## Evaluation protocol

Each item is answered by: (1) watching the video and locating the moments relevant to
the question; (2) extracting observable search anchors (a landmark, on-screen text,
logo, spoken name, distinctive object, …); (3) searching the open web from those
anchors and keeping the URL + a supporting verbatim quote for every external fact;
(4) combining video observations with the retrieved documents. Answers should be the
shortest response that fully satisfies the question; for multi-part questions, answer
each part semicolon-separated in the order asked, and write `UNKNOWN` for any part that
cannot be established rather than guessing. `audio_required` / `image_search_required`
flag items that additionally need the audio track or image-based search.

## Limitations and notes

- **No ground-truth is released.** The public items are sanitized for open evaluation;
  do not source answers from any other file in this package.
- Questions require live open-web access; retrievable facts may change over time.
- Videos are sourced externally. **Licensing/usage of the underlying video content is
  the responsibility of the dataset owner and downstream users** — please confirm the
  terms before redistribution or commercial use. (License field intentionally left
  unset pending the owner's confirmation.)

## Citation

```bibtex
@misc{vdr_benchmark,
  title  = {VDR: Video Deep Research Benchmark},
  author = {hh721},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/hh721/vdr_temp}}
}
```