File size: 5,807 Bytes
51bbc4b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
license: cc
pretty_name: AgentVidBench (Sample)
task_categories:
  - video-text-to-text
  - multiple-choice
  - question-answering
configs:
  - config_name: questions
    data_files: questions.jsonl
  - config_name: videos
    data_files: videos.jsonl
---

# AgentVidBench (Sample): A Multi-Hop Video Question Answering Benchmark for Evaluating MLLM Agents

This repository is a **representative sample** of [AgentVidBench](https://huggingface.co/datasets/agentvidbench/agentvidbench), provided so reviewers can inspect data quality without downloading the full ~4GB+ corpus. The full dataset remains available at the link above.

<div align="center">
  <img src="assets/example_q56.png" alt="Q56 — Bicep Curls Before &quot;One More&quot; (example task with human-curated reasoning trajectory; question shown for illustration only and is not included in this sample)" width="75%">
</div>

## Sample selection

The sample contains the **first 10 questions** (`question_id` 1–10) and the **9 unique videos** they reference. IDs and filenames are preserved from the full dataset — no renumbering — so any record in this sample can be cross-referenced with the full dataset by its `question_id` or `file_name`.

- `question_id` 1–10 (contiguous)
- Videos referenced: `video1`, `video2`, `video4`, `video5`, `video6`, `video7`, `video8`, `video9`, `video10` (9 unique; `video2` is referenced by both questions 2 and 3, and `video3` is not referenced by any of the first 10 questions)
- Whisper transcripts (`.srt`) are included for each of the 9 videos
- All schemas, fields, and file formats are identical to the full dataset

## Layout

```
.
├── README.md
├── questions.jsonl        # 10 rows — first 10 questions of the full dataset
├── videos.jsonl           # 9 rows — videos referenced by those questions
├── videos/
│   └── video*.mp4         # 9 video files
└── transcripts/
    └── video*.srt         # 9 Whisper transcripts
```

## `questions.jsonl` schema

| Field                | Type                | Notes                                                         |
|----------------------|---------------------|---------------------------------------------------------------|
| `question_id`        | int                 | 1–100 in the full dataset (1–10 in this sample)               |
| `video_path`         | str                 | relative path to the video file (`videos/videoN.mp4`); equals the `file_name` in `videos.jsonl` |
| `transcript_path`    | str                 | relative path to the Whisper transcript (`transcripts/videoN.srt`); shares the basename with `video_path` |
| `title`              | str                 | short descriptive title                                       |
| `question_text`      | str                 | full question prompt                                          |
| `options`            | list[str] (len 26)  | answer choices in order A, B, C, … Z                          |
| `answer`             | str                 | correct answer letter, A–Z                                    |
| `answer_explanation` | str                 | reasoning + reference for the answer                          |
| `skills`             | list[str]           | core capabilities the question tests                          |
| `difficulty`         | str                 | `easy` / `medium` / `hard` / `very_hard`                      |
| `options_type`       | str                 | distractor strategy                                           |
| `categories`         | list[str]           | cognitive task tags                                           |
| `milestones`         | list[step]          | sub-tasks the question tests; each step is `{id, type, description}` |
| `ecom_required`      | bool                | true iff the answer cannot be locked in until later evidence in the video has been observed (i.e., the full clip must be watched) |
| `trajectory`         | dict                | Reasoning trace. Shape: `{steps: list[step], final_answer: str, model: str, temperature: float, generated_at: str, elapsed_seconds: float}`. Each step is `{tool: str, args: dict, thought: str, observation: str}`; `args` is heterogeneous per tool (HF infers it as a `Json` feature). Populated for all questions. |

## `videos.jsonl` schema

| Field              | Type           | Notes                                              |
|--------------------|----------------|----------------------------------------------------|
| `file_name`        | str            | relative path to the video file (`videos/videoN.mp4`) |
| `title`            | str            | —                                                  |
| `duration`         | str            | `M:SS`                                             |
| `genre`            | str            | one of 12 enum values                              |
| `production_style` | list[str]      | shot/edit style tags                               |
| `audio_type`       | list[str]      | audio characteristics                              |
| `length_category`  | str            | `short` / `medium` / `long`                        |
| `url`              | str            | source URL                                         |
| `license`          | str            | license                                            |
| `source`           | str            | —                                                  |
| `author`           | str            | —                                                  |

## Download

```bash
hf download agentvidbench/agentvidbench-sample --repo-type dataset --local-dir dataset-sample
```

The full dataset is at [`agentvidbench/agentvidbench`](https://huggingface.co/datasets/agentvidbench/agentvidbench).