File size: 8,649 Bytes
a08d02b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5aafb9e
a08d02b
5aafb9e
a08d02b
5aafb9e
a08d02b
5aafb9e
a08d02b
5aafb9e
 
 
 
 
 
 
a08d02b
5aafb9e
 
 
 
 
 
 
a08d02b
5aafb9e
 
 
 
 
 
 
 
ec1e768
 
 
 
 
 
 
a08d02b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ec1e768
a08d02b
 
 
 
 
 
 
 
 
 
 
 
 
daa1023
 
 
 
 
 
a08d02b
ec1e768
 
 
 
 
 
 
 
 
 
 
 
a08d02b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ec1e768
 
 
a08d02b
 
 
 
 
 
 
 
 
 
 
2dce326
950cd58
a08d02b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
---
pretty_name: Ko-MultimodalMind2Web
license: cc-by-4.0
language:
- ko
task_categories:
- image-text-to-text
size_categories:
- 10K<n<100K
tags:
- korean
- gui-agent
- web-agent
- human-annotated
- evaluation
modalities:
- image
- text
configs:
- config_name: default
  data_files:
  - split: test
    path: examples/test.parquet
---

# 🇰🇷 Ko-MultimodalMind2Web

An offline Korean web-agent benchmark, built manually by native Korean annotators.

📋 549 tasks · 🖼️ 12,905 scored steps · 🌐 144 websites · ✍️ ~480 annotator-hours · 🧭 Offline evaluation

## 📌 Overview

Ko-MultimodalMind2Web is an offline web-agent benchmark for Korean websites.
It contains 549 tasks and 12,905 scored action steps across
144 websites, collected from human browser demonstrations. The published
task interface follows [Multimodal Mind2Web](https://huggingface.co/datasets/osunlp/Multimodal-Mind2Web):
given a Korean task, the current screenshot, and preceding action history, an
agent predicts the next operation, target, and action value. Evaluation reports
element accuracy, step success, and task success.

## ✍️ Built by Human Annotators

All tasks and reference demonstrations were created manually by Korean human
annotators. They designed realistic requests, operated the websites themselves,
and recorded browser demonstrations of each task.

| Item | Details |
| --- | --- |
| 👤 Annotators | 2 native Korean speakers |
| ⏱️ Total effort | About 480 annotator-hours |

This is the shared collection and annotation effort behind all three Korean
releases in the suite below. The estimate is based on two annotators working
four hours per day, five days per week, for approximately three months
(about twelve weeks).

The benchmark follows the offline task/trajectory/action organization used by
[Mind2Web](https://arxiv.org/abs/2306.06070) and
[Multimodal Mind2Web / SeeAct](https://huggingface.co/datasets/osunlp/Multimodal-Mind2Web):
language instructions and screenshots are paired with next-action targets.
The Korean task demonstrations and annotations were collected manually.
The released operation inventory, action history, and accepted alternatives
are explicitly retained so evaluation is reproducible on this collection.

## How the benchmark was constructed and verified

1. **Human task design and demonstration.** Korean annotators created realistic
   tasks on Korean websites and Korean-language interfaces, then performed the
   complete tasks themselves. Browser events, screenshots, and target-element
   information were captured from their actual interactions.
2. **Step-level annotation.** The recorded demonstrations were organized into
   task-linked steps with an operation, argument/value, target geometry when
   relevant, and the preceding action history. Both original and concise
   task-instruction variants are supplied.
3. **Manual target correction.** Native browser dialogs, separate popup windows,
   operating-system select menus, and dynamically replaced controls were
   inspected using their screenshots. Missing or incorrectly aligned targets
   were corrected manually. Selection actions were checked against subsequent
   screenshots to distinguish the selected value from a transient highlight.
4. **Alternative-answer review.** Equivalent form-submission actions were
   represented as alternative valid answers. The final release includes
   **350 manually located alternative CLICK targets**, together with
   template-matched alternatives and valid Enter-key submissions. Operation,
   value, and location must belong to the same accepted answer candidate.
5. **Consistency and scoring checks.** All 12,905 scored-step IDs are unique;
   all steps resolve to one of the 549 tasks and agree with task-level step
   counts. All 7,197 primary target boxes are finite, have positive area, and
   lie within their screenshot dimensions. Gold-label round-trip checks return
   1.0 for element accuracy, step success, and task success using the included
   scorer.

## 📊 Dataset statistics

### Recorded operation distribution

| Category | Count | Share |
| --- | --- | --- |
| CLICK | 5,879 | 45.56% |
| TYPE | 1,399 | 10.84% |
| HOTKEY | 444 | 3.44% |
| SCROLL | 5,007 | 38.80% |
| CLEAR | 15 | 0.12% |
| GO_BACK | 134 | 1.04% |
| DRAG | 27 | 0.21% |

The 549 human reference trajectories contain 14,042 recorded steps. Offline
next-action scoring is restricted to the seven interaction operations listed
above. URL navigation (`GOTO`), waiting (`WAIT`), and task termination (`FINISH`)
are excluded from step-level scoring, leaving 12,905 scored interaction steps.
`statistics.json` includes per-website counts, task categories, trajectory-length
statistics, and alternative-action counts.

## 🧩 KORMo-VL Benchmark Suite

This dataset is part of the KORMo-VL Korean GUI-agent benchmark suite. Each
release is built from manually collected Korean data and follows the protocol
of an established international benchmark.

| Dataset | Setting | Follows |
| --- | --- | --- |
| [Ko-Screenspot](https://huggingface.co/datasets/KORMo-VL/Ko-Screenspot) | Single-image GUI grounding | [ScreenSpot](https://aclanthology.org/2024.acl-long.505/) |
| [Ko-MultimodalMind2Web](https://huggingface.co/datasets/KORMo-VL/Ko-MultimodalMind2Web) | Offline next-action prediction | [Multimodal Mind2Web](https://huggingface.co/datasets/osunlp/Multimodal-Mind2Web) |
| [Ko-Online-Mind2Web](https://huggingface.co/datasets/KORMo-VL/Ko-Online-Mind2Web) | Live task completion | [Online Mind2Web](https://github.com/OSU-NLP-Group/Online-Mind2Web) |

## Download and preview

This repository is private. Authenticate with a Hugging Face account that has
access to the repository.

```python
from huggingface_hub import snapshot_download

snapshot_download(
    "KORMo-VL/Ko-MultimodalMind2Web",
    repo_type="dataset",
    local_dir="Ko-MultimodalMind2Web",
    allow_patterns=["README.md", "statistics.json", "evaluate.py", "multimodal_mind2web_ko.tar.gz"],
    token=True,
)
```

The dataset viewer contains **10 illustrative examples**. The complete benchmark
is provided in the separate full-data file listed above. These examples are
selected from the full data.

```python
from datasets import load_dataset

preview = load_dataset("KORMo-VL/Ko-MultimodalMind2Web", split="test", token=True)
```

Extract the full data from the downloaded repository directory:

```bash
cd Ko-MultimodalMind2Web
tar -xzf multimodal_mind2web_ko.tar.gz
```

## Evaluation

Use the included `evaluate.py` to report element accuracy, step success rate,
and task success rate on the released scored steps. The reference scorer
supports accepted alternative actions.

## Contact

* KyungTae Lim, Professor at KAIST. `ktlim@kaist.ac.kr`

## Contributor

* Yerin Nam ([yerinnam@kaist.ac.kr](mailto:yerinnam@kaist.ac.kr))
* Minkyung Cho ([minkyung.cho@kaist.ac.kr](mailto:minkyung.cho@kaist.ac.kr))
* Minkyu Kim ([minkyukim@kaist.ac.kr](mailto:minkyukim@kaist.ac.kr))
* Seungwoo Song ([sswoo@kaist.ac.kr](mailto:sswoo@kaist.ac.kr))
* Hangyeol Yoo ([hg.yoo@kaist.ac.kr](mailto:hgyoo@kaist.ac.kr))
* Gyun Lee ([gyunini@kaist.ac.kr](mailto:gyunini@kaist.ac.kr))
* KyungTae Lim ([ktlim@kaist.ac.kr](mailto:ktlim@kaist.ac.kr))

## Citation

References for KORMo and the benchmark protocol followed by this release:

```bibtex
@misc{KORMo,
  author = {Minjun Kim and Hyeonseok Lim and Hangyeol Yoo and Inho Won and Seungwoo Song and Minkyung Cho and Junghun Yuk and Changsu Choi and Dongjae Shin and Huije Lee and Hoyun Song and Alice Oh and KyungTae Lim},
  title = {KORMo: Korean Open Reasoning Model for Everyone},
  year = {2025},
  publisher = {GitHub},
  journal = {Technical Report},
  url = {https://arxiv.org/abs/2510.09426}
}
```

```bibtex
@inproceedings{zheng2024seeact,
  title = {GPT-4V(ision) is a Generalist Web Agent, if Grounded},
  author = {Boyuan Zheng and Boyu Gou and Jihyung Kil and Huan Sun and Yu Su},
  booktitle = {Forty-first International Conference on Machine Learning},
  year = {2024},
  url = {https://openreview.net/forum?id=piecKJ2DlB}
}
```

```bibtex
@inproceedings{deng2023mind2web,
  title = {Mind2Web: Towards a Generalist Agent for the Web},
  author = {Xiang Deng and Yu Gu and Boyuan Zheng and Shijie Chen and Sam Stevens and Boshi Wang and Huan Sun and Yu Su},
  booktitle = {Advances in Neural Information Processing Systems},
  volume = {36},
  pages = {28091--28114},
  year = {2023},
  publisher = {Curran Associates, Inc.},
  url = {https://proceedings.neurips.cc/paper_files/paper/2023/file/5950bf290a1570ea401bf98882128160-Paper-Datasets_and_Benchmarks.pdf}
}
```