---
license: cc-by-4.0
language:
- ko
task_categories:
- image-text-to-text
size_categories:
- 10K<n<100K
tags:
- gui-agent
- grounding
- korean
- web-agent
- evaluation
configs:
- config_name: screenspot_ko
  data_files:
  - split: test
    path: screenspot_ko/examples/test.parquet
- config_name: multimodal_mind2web_ko
  data_files:
  - split: test
    path: multimodal_mind2web_ko/examples/test.parquet
- config_name: online_mind2web_ko
  data_files:
  - split: test
    path: online_mind2web_ko/examples/test.parquet
---

# Ko-Bench: Korean Web GUI Evaluation

**Ko-Bench was created by human annotators.**

The suite brings together three Korean benchmarks for GUI grounding, offline
action prediction, and live website task completion. The Hub viewer contains
**10 examples per benchmark**. Complete image benchmarks are provided in gzip
archives; Online Mind2Web is provided as an uncompressed JSON file.
The examples are subsets of the full release, not additional evaluation tasks.

| Benchmark | Complete release | Viewer examples | Full download |
| --- | --- | ---: | --- |
| Korean ScreenSpot | 3,319 grounding questions | 10 | [screenspot_ko.tar.gz](screenspot_ko/screenspot_ko.tar.gz) |
| Korean Multimodal Mind2Web | 12,905 scored steps from 549 tasks | 10 steps | [multimodal_mind2web_ko.tar.gz](multimodal_mind2web_ko/multimodal_mind2web_ko.tar.gz) |
| Korean Online Mind2Web | 312 live website tasks | 10 | [online_mind2web_ko_tasks.json](online_mind2web_ko/online_mind2web_ko_tasks.json) |

## Download the complete benchmarks

```python
from huggingface_hub import snapshot_download

snapshot_download(
    "KORMo-VL/Ko-Bench",
    repo_type="dataset",
    local_dir="Ko-Bench",
    allow_patterns=["README.md", "*/*.tar.gz", "online_mind2web_ko/*.json"],
)
```

Extract from the release root:

```bash
cd Ko-Bench
tar -xzf screenspot_ko/screenspot_ko.tar.gz
tar -xzf multimodal_mind2web_ko/multimodal_mind2web_ko.tar.gz
```

After extraction, ScreenSpot has `screenspot_ko/samples.jsonl` and `images/`;
Multimodal Mind2Web has `multimodal_mind2web_ko/steps.jsonl`, `tasks.jsonl`,
`tasks_rewrite.jsonl`, and `images/`. Each JSONL `image` path is relative to
its benchmark directory. Online Mind2Web is a JSON array of task records with exactly five fields:
`task_id`, `instruction`, `start_url`, `task_level`, and `task_type`.
`task_type` is a list of labels; its values are preserved from the original task types.

## Load the 10-example previews

```python
from datasets import load_dataset

preview = load_dataset("KORMo-VL/Ko-Bench", "screenspot_ko", split="test")
# Other configurations: multimodal_mind2web_ko, online_mind2web_ko
```

The two image previews embed the original image bytes directly in Parquet, so
the viewer does not need to download the full image archives. Complete example
records are provided in `examples/samples.jsonl`. The offline Mind2Web preview
also exposes `action_json` and `previous_actions`. `task_id` is the final column
in both image previews. Matching unmodified PNG files are included.
Preview selection is illustrative and is not intended for reporting benchmark scores.

## Tasks and evaluation

- **ScreenSpot:** locate the UI element described by a Korean instruction.
  A predicted point is correct when it falls inside `gold_bbox`.
- **Multimodal Mind2Web:** predict the next action from a screenshot, task
  instruction, and action history. Evaluate element accuracy, step success,
  and task success. There are 7,580 scored steps with a target box.
  The archives contain the 12,905 scored steps; the original complete
  trajectories contain 14,042 steps, including unscored `GOTO`, `WAIT`, and `FINISH`.
- **Online Mind2Web:** execute tasks on live websites and assess whether
  the requested result was achieved. The release excludes 14 date-dependent
  tasks from the previous 326-task no-login set. It includes `task_level`
  and potentially overlapping `task_type` labels. Website content can change.

Pixel coordinates and normalized `*_norm` fields retain their recorded
conventions. No action vocabulary or coordinate conversion is applied in this release.

## About KORMo-VL

KORMo-VL is a vision-language model developed from scratch by the
[KAIST MLP Lab](https://sites.google.com/view/aailab), built on top of KORMo-10B.
This dataset release provides Korean web GUI evaluation resources under the
[KORMo-VL organization](https://huggingface.co/KORMo-VL).

## 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 ([hgyoo@kaist.ac.kr](mailto:hgyoo@kaist.ac.kr))
* KyungTae Lim ([ktlim@kaist.ac.kr](mailto:ktlim@kaist.ac.kr))

## Citation

References for KORMo and the original benchmark formats are provided below.

### KORMo

```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}
}
```

### ScreenSpot

[SeeClick: Harnessing GUI Grounding for Advanced Visual GUI Agents](https://aclanthology.org/2024.acl-long.505/)

```bibtex
@inproceedings{cheng2024seeclick,
  title = {SeeClick: Harnessing GUI Grounding for Advanced Visual GUI Agents},
  author = {Kanzhi Cheng and Qiushi Sun and Yougang Chu and Fangzhi Xu and Yantao Li and Jianbing Zhang and Zhiyong Wu},
  booktitle = {Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2024},
  pages = {9313--9332},
  publisher = {Association for Computational Linguistics},
  doi = {10.18653/v1/2024.acl-long.505},
  url = {https://aclanthology.org/2024.acl-long.505/}
}
```

### Multimodal Mind2Web

The [official Multimodal Mind2Web dataset card](https://huggingface.co/datasets/osunlp/Multimodal-Mind2Web) cites SeeAct and the original Mind2Web work.

```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}
}
```

### Online Mind2Web

[An Illusion of Progress? Assessing the Current State of Web Agents](https://github.com/OSU-NLP-Group/Online-Mind2Web)

```bibtex
@inproceedings{xue2025an,
  title = {An Illusion of Progress? Assessing the Current State of Web Agents},
  author = {Tianci Xue and Weijian Qi and Tianneng Shi and Chan Hee Song and Boyu Gou and Dawn Song and Huan Sun and Yu Su},
  booktitle = {Second Conference on Language Modeling},
  year = {2025},
  url = {https://openreview.net/forum?id=6jZi4HSs6o}
}
```

### Mind2Web

Both Multimodal Mind2Web and Online Mind2Web also request citation of the original Mind2Web work.

```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}
}
```