Hugging Face's logo

Datasets:
longvideoagent
/
LongTVQA_plus

Modalities:
Text
Formats:
json
Languages:
English
Size:
10K - 100K
ArXiv:
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
Community
1
LongTVQA_plus
File size: 4,846 Bytes
2e9afd8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ab96ef4
 
 
 
 
2e9afd8
 
 
 
 
 
 
 
 
 
 
 
 
ab96ef4
 
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
123
124
125
126
127
128
129
130
131
132
133
134
135
 
---
license: mit
language:
- en
pretty_name: LongTVQA+
---

# LongTVQA+ Dataset

This repository contains the **LongTVQA+** dataset in JSON format.

LongTVQA+ is built upon the original **TVQA+** dataset, with the key difference that it **extends the question grounding scope from short clip-level segments (≈1 minute) to long episode-level videos (up to ~20 minutes)**.  
This enables research on long-form video understanding, long-range temporal reasoning, and fine-grained spatio-temporal grounding in realistic TV show episodes.

In addition to the extended temporal scope, LongTVQA+ preserves and leverages the rich annotations provided in TVQA+, including:


1. Frame-level bounding box annotations for visual concept words appearing in questions and correct answers.
2. Refined timestamp annotations aligned with long episode-level context.

Please refer to the original **TVQA+ paper** for details on the annotation protocol and baseline evaluations.

---

## Files

- `LongTVQA_plus_train.json` — training split (23,545 QA samples)
- `LongTVQA_plus_val.json` — validation split (3,017 QA samples)
- `LongTVQA_plus_subtitle_clip_level.json` — clip-level subtitles indexed by video clip (4,198 clips)
- `LongTVQA_plus_subtitle_episode_level.json` — episode-level subtitles indexed by episode (220 episodes)

---

## QA JSON Format

Each entry in `LongTVQA_plus_train.json` and `LongTVQA_plus_val.json` is a dictionary with the following fields:

| Key | Type | Description |
| --- | --- | --- |
| `qid` | int | Question ID (same as in TVQA+). |
| `q` | str | Question text. |
| `a0` ... `a4` | str | Five multiple-choice answers. |
| `answer` | str | Correct answer key (`"a0"`–`"a4"`). |
| `ts` | list | Refined timestamp annotation. For example, `[0, 5.4]` indicates the localized temporal span starts at 0s and ends at 5.4s. |
| `episode_name` | str | Episode ID (e.g. `s01e02`). |
| `occur_clip` | str | Video clip name. Format: `{show_name_abbr}_s{season}e{episode}_seg{segment}_clip_{clip}`. Episodes are typically divided into two segments separated by the opening theme. For **The Big Bang Theory**, `{show_name_abbr}` is omitted (e.g. `s05e02_seg02_clip_00`). |
| `bbox` | dict | Frame-level bounding box annotations sampled at 3 FPS. Keys are frame indices. Values are lists of bounding boxes with `img_id`, `top`, `left`, `width`, `height`, and `label`. |

---

### QA Sample

```json
{
  "answer": "a1",
  "qid": 134094,
  "ts": [5.99, 11.98],
  "a1": "Howard is talking to Raj and Leonard",
  "a0": "Howard is talking to Bernadette",
  "a3": "Howard is talking to Leonard and Penny",
  "a2": "Howard is talking to Sheldon , and Raj",
  "q": "Who is Howard talking to when he is in the lab room ?",
  "episode_name": "s05e02",
  "occur_clip": "s05e02_seg02_clip_00",
  "a4": "Howard is talking to Penny and Bernadette",
  "bbox": {
    "14": [
      {
        "img_id": 14,
        "top": 153,
        "label": "Howard",
        "width": 180,
        "height": 207,
        "left": 339
      },
      {
        "img_id": 14,
        "top": 6,
        "label": "lab",
        "width": 637,
        "height": 354,
        "left": 3
      }
    ],
    "20": [],
    "26": [],
    "32": [],
    "38": []
  }
}
```

---

## Subtitles JSON Format

Two subtitle files are provided to support different temporal granularities:

| File                                        | Key            | Type | Description                                                                                                         |
| ------------------------------------------- | -------------- | ---- | ------------------------------------------------------------------------------------------------------------------- |
| `LongTVQA_plus_subtitle_clip_level.json`    | `vid_name`     | str  | Clip-level subtitle text, with utterances separated by `<eos>`.                                                     |
| `LongTVQA_plus_subtitle_episode_level.json` | `episode_name` | str  | Episode-level subtitle text, including clip markers such as `<seg01_clip_00>`, and utterances separated by `<eos>`. |

---

### Subtitles Sample

```json
{
  "s09e14_seg02_clip_04": "Sheldon : That 's a risk I'm willing to take ! <eos> Amy : Well , this is so nice . <eos> ..."
}
```

---




## License

This dataset is released under the **MIT License**.
## 📝 Citation

If you find our work helpful, please cite:

```bibtex
@misc{liu2025longvideoagentmultiagentreasoninglong,
      title={LongVideoAgent: Multi-Agent Reasoning with Long Videos}, 
      author={Runtao Liu and Ziyi Liu and Jiaqi Tang and Yue Ma and Renjie Pi and Jipeng Zhang and Qifeng Chen},
      year={2025},
      eprint={2512.20618},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={[https://arxiv.org/abs/2512.20618](https://arxiv.org/abs/2512.20618)}, 
}