Datasets:
sample_id stringclasses 10
values | image imagewidth (px) 1.28k 1.28k | instruction stringclasses 9
values | website stringclasses 10
values | gold_bbox listlengths 4 4 | element_type stringclasses 2
values | task_id stringclasses 10
values |
|---|---|---|---|---|---|---|
screenspot::0714_X_2_1#2 | 이메일 또는 사용자 이름을 입력해 보세요. | X_2 | [
51.5,
619.5,
601.4,
706.8
] | text | 0714_X_2_1 | |
screenspot::0804_amoremall_1#9 | 베스트 판매 상품을 확인해줘. | amoremall | [
190.2,
117.7,
236.5,
162.2
] | text | 0804_amoremall_1 | |
screenspot::0714_ebsi_1#2 | 팝업 창을 닫아줘. | ebsi | [
767.9,
644.2,
824.7,
665.3
] | text | 0714_ebsi_1 | |
screenspot::0714_google_docs_1#2 | 이메일 또는 휴대전화로 로그인해 보세요. | google_docs | [
677.2,
151.7,
1231.4,
226.2
] | icon_widget | 0714_google_docs_1 | |
screenspot::0714_instagram_2_1#14 | 정보 저장 창을 닫아줘. | instagram_2 | [
379.8,
471.2,
878.2,
497
] | text | 0714_instagram_2_1 | |
screenspot::0714_naver_webtoon_1#2 | 네이버 계정으로 로그인해. | naver_webtoon | [
1063.4,
17.5,
1101.9,
34.2
] | text | 0714_naver_webtoon_1 | |
screenspot::0714_ohou_se_1#4 | 네이버 아이디로 로그인해줘. | ohou_se | [
357.4,
517.2,
900.6,
548.8
] | icon_widget | 0714_ohou_se_1 | |
screenspot::0714_pexels_1#9 | 쿠키 사용을 모두 허용하자. | pexels | [
630.7,
723.9,
744.3,
765.6
] | text | 0714_pexels_1 | |
screenspot::0714_pinterest_1#2 | 네이버 계정으로 로그인해 주세요. | pinterest | [
984.5,
22.9,
1093.6,
91.7
] | text | 0714_pinterest_1 | |
screenspot::0715_cgv_1#3 | 팝업 창을 닫아줘. | cgv | [
844.1,
730.9,
877.9,
754.4
] | text | 0715_cgv_1 |
🇰🇷 Ko-Screenspot
A Korean GUI grounding benchmark, built manually by native Korean annotators.
📋 3,319 samples · 🌐 142 websites · ✍️ ~480 annotator-hours · 🎯 GUI grounding
📌 Overview
Ko-Screenspot is a single-image GUI grounding benchmark for Korean websites. It contains 3,319 screenshot–instruction pairs across 142 websites, derived from 540 human-recorded tasks. Instructions are written in Korean, and each sample includes an annotated target element. The published task interface follows ScreenSpot: given a screenshot and a grounding instruction, an agent predicts the location of the requested UI element.
✍️ 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 task interface follows ScreenSpot / SeeClick: one screenshot, one grounding instruction, and a target element evaluated by point-in-box accuracy. The Korean screenshots and demonstrations come from the manual collection described above. The task IDs preserve the connection to their source human demonstrations.
How the benchmark was constructed and verified
- Manual Korean task collection. Korean annotators designed realistic website tasks and performed them in a browser. Screenshots, actions, and target-element information were recorded from these human demonstrations.
- Grounding annotation. Element-targeting steps were converted into screenshot–instruction–target examples. Visible Korean labels and accessible element names were used to anchor the requested UI function; instructions were curated into concise Korean referring expressions.
- Annotation normalization. Target geometry was aligned to the actual screenshot dimensions, and the reference point was set to the box center. Repeated targets within a demonstration, unusable labels, and oversized page-container boxes were filtered. Non-Korean interfaces were screened out.
- Manual ambiguity review. Human reviewers examined 267 potentially problematic questions, inspecting the screenshots and target annotations. 163 questions were excluded because of multiple plausible targets or misaligned labels, and 104 were retained. The resulting release contains 3,319 questions.
- Release consistency verification. The released questions have unique sample IDs; all 3,319 target boxes are finite, have positive area, and lie within their screenshot dimensions. Image references and archive contents were checked. A gold-label round-trip through the included scorer returns 1.0 grounding accuracy, verifying data/scorer consistency.
📊 Dataset statistics
Target element types
| Category | Count | Share |
|---|---|---|
| text | 2,874 | 86.59% |
| icon_widget | 445 | 13.41% |
🧩 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 | Single-image GUI grounding | ScreenSpot |
| Ko-MultimodalMind2Web | Offline next-action prediction | Multimodal Mind2Web |
| Ko-Online-Mind2Web | Live task completion | Online Mind2Web |
Download and preview
This repository is private. Authenticate with a Hugging Face account that has access to the repository.
from huggingface_hub import snapshot_download
snapshot_download(
"KORMo-VL/Ko-Screenspot",
repo_type="dataset",
local_dir="Ko-Screenspot",
allow_patterns=["README.md", "statistics.json", "evaluate.py", "screenspot_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.
from datasets import load_dataset
preview = load_dataset("KORMo-VL/Ko-Screenspot", split="test", token=True)
Extract the full data from the downloaded repository directory:
cd Ko-Screenspot
tar -xzf screenspot_ko.tar.gz
Evaluation
Use the included evaluate.py to measure grounding accuracy: a prediction is
correct when its point falls inside the reference target box. Report overall
accuracy and the text/icon-widget breakdown.
Contact
- KyungTae Lim, Professor at KAIST.
ktlim@kaist.ac.kr
Contributor
- Yerin Nam (yerinnam@kaist.ac.kr)
- Minkyung Cho (minkyung.cho@kaist.ac.kr)
- Minkyu Kim (minkyukim@kaist.ac.kr)
- Seungwoo Song (sswoo@kaist.ac.kr)
- Hangyeol Yoo (hgyoo@kaist.ac.kr)
- Gyun Lee (gyunini@kaist.ac.kr)
- KyungTae Lim (ktlim@kaist.ac.kr)
Citation
References for KORMo and the benchmark protocol followed by this release:
@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}
}
@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/}
}
- Downloads last month
- -