Dataset Viewer
Auto-converted to Parquet Duplicate
task_id
stringclasses
10 values
instruction
stringclasses
10 values
start_url
stringclasses
10 values
task_level
stringclasses
3 values
task_type
listlengths
1
3
easy_final_data_0715/local/google_maps_1
μ΄λ²ˆμ— μˆ˜μ›μ— λ†€λŸ¬κ°€λŠ”λ° μˆ˜μ›μ—­ κ·Όμ²˜μ—μ„œ λ°₯을 λ¨Ήμ–΄μ•Ό ν•΄. ꡬ글 λ§΅μŠ€μ—μ„œ μˆ˜μ›μ—­ 근처 맛집을 검색 ν›„ 평점 4.5점 이상인 식당쀑 κ°€μž₯ μœ„μ— λ‚˜μ˜€λŠ” κ°€κ²Œ 이름 μ•Œλ €μ€˜.
https://www.google.co.kr/maps/?hl=ko&entry=ttu&g_ep=EgoyMDI2MDcxMi4wIKXMDSoASAFQAw%3D%3D
easy
[ "filter", "ranking", "site_interaction" ]
easy_final_data_0715/local/daangn_2
돈이 ν•„μš”ν•΄μ„œ μ•„λ₯΄λ°”μ΄νŠΈ 일자리λ₯Ό μ°Ύκ³  μžˆμ–΄. λ‹Ήκ·Όλ§ˆμΌ“μ—μ„œ μ„œμšΈμ‹œ μ„œμ΄ˆκ΅¬ λ°©λ°°1동에 μžˆλŠ” 주방보쑰/μ„€κ±°μ§€ 곡고 쀑 평일에 κ°€λŠ₯ν•œ μ•„λ₯΄λ°”μ΄νŠΈ ν•œ κ³³ 곡고글 λ³΄μ—¬μ€˜. λ§Œμ•½ λ°©λ°°1동에 μ—†μœΌλ©΄ λ°©λ°°2λ™μœΌλ‘œ κ²€μƒ‰ν•΄μ€˜.
https://www.daangn.com/kr/
medium
[ "filter", "return_navigation" ]
easy_final_data_0714/local/ohou_se_2
거싀에 μ†ŒνŒŒκ°€ μ—†μ–΄μ„œ μ†ŒνŒŒλ₯Ό 사야해. 였늘의 μ§‘μ—μ„œ 였늘좜발 κ°€λŠ₯ν•œ 2인 μ†ŒνŒŒμ˜ 첫번째 μƒν’ˆλͺ…을 μ•Œλ €μ€˜ 리뷰 λ§Žμ€μˆœμœΌλ‘œ 가격은 10λ§Œμ›μ—μ„œ 20λ§Œμ› μ‚¬μ΄λ‘œ.
https://ohou.se/
hard
[ "filter", "ranking" ]
easy_final_data_0715/local/cgv_4
주말에 혼자 μ—¬μœ λ‘­κ²Œ μ˜ν™”λ₯Ό 보렀고 ν•΄. cgv μ‚¬μ΄νŠΈμ—μ„œ μƒμ˜μ€‘μΈ μ˜ν™” 쀑 예맀율 1μœ„μΈ μ˜ν™”λ₯Ό ν™•μΈν•΄μ€˜. 그리고 λ‹€μŒμ£Ό ν† μš”μΌμ— λŒ€μ „cgvμ—μ„œ μƒμ˜ν•˜λŠ” ν•΄λ‹Ή μ˜ν™”μ˜ μ‹œκ°„ν‘œλ₯Ό μ‘°νšŒν•΄μ„œ, 22μ‹œ 이후 μƒμ˜ μ˜ˆμ •μΈ κ΄€μ˜ μƒμ˜ μ‹œκ°„λŒ€μ™€ 이름을 λͺ¨λ‘ μ•Œλ €μ€˜.
https://cgv.co.kr/
hard
[ "filter", "ranking", "return_navigation" ]
easy_final_data_0715/local/dict_naver_1
λΆ€μ§€κΈ°μˆ˜λΌλŠ” 단어가 무슨 말인지 λ„μ €νžˆ 이해가 μ•ˆκ°€. 넀이버 사전에 λΆ€μ§€κΈ°μˆ˜λ₯Ό 검색해주고 무슨 λœ»μΈμ§€ μ•Œλ €μ€˜.
https://dict.naver.com/
medium
[ "information_lookup" ]
easy_final_data_0715/local/ebsi_1
여름방학 λ™μ•ˆ 2μ£Ό 정도 μ‹œκ°„μ„ λ‚΄μ„œ μˆ˜ν•™β…  κ°œλ…μ„ μ²˜μŒλΆ€ν„° λ‹€μ‹œ μ •λ¦¬ν•˜λ €κ³  ν•©λ‹ˆλ‹€. EBSiμ—μ„œ μˆ˜ν•™β…  κ°•μ’Œλ₯Ό 찾아보고 기초 κ°œλ…λΆ€ν„° μ„€λͺ…ν•˜λŠ” κ°•μ˜ 쀑 ν•˜λ‚˜λ₯Ό μ„ νƒν•΄μ£Όμ„Έμš”. κ°•μ’Œ 상세 νŽ˜μ΄μ§€μ—μ„œ μ„ μƒλ‹˜ 이름과 전체 κ°•μ˜ 수, 첫 번째 κ°•μ˜μ—μ„œ λ‹€λ£¨λŠ” λ‚΄μš©μ„ ν™•μΈν•΄μ£Όμ„Έμš”
https://www.ebsi.co.kr/ebs/pot/poti/main.ebs
hard
[ "information_lookup" ]
easy_final_data_0715/local/hogang_nono_1
뢀동산에 μš”μ¦˜ 관심이 μžˆμ–΄μ„œ λ³΄λŠ” 쀑이야. ν˜Έκ°±λ…Έλ…Έ μ‚¬μ΄νŠΈμ—μ„œ λŒ€μ „ μœ μ„±κ΅¬μ˜ 졜근 평당가와 κ±°λž˜λŸ‰μ΄ κ°€μž₯ λ§Žμ€ μ•„νŒŒνŠΈ 이름 μ•Œλ €μ€˜.
https://hogangnono.com/
medium
[ "filter", "ranking" ]
easy_final_data_0715/local/mcdonalds_1
햄버거λ₯Ό λ¨Ήκ³  싢은데 λ‚΄κ°€ μƒˆμš° μ•Œλ ˆλ₯΄κΈ°κ°€ μžˆμ–΄. λ§₯λ„λ‚ λ“œ μ‚¬μ΄νŠΈμ—μ„œ μŠˆλΉ„ 버거 μ•Œλ ˆλ₯΄κΈ° 정보λ₯Ό μ°Ύμ•„μ£Όκ³  μƒˆμš°κ°€ μžˆλŠ”μ§€ ν™•μΈν•΄μ„œ λ§ν•΄μ€˜.
https://www.mcdonalds.co.kr/kor/
easy
[ "information_lookup" ]
easy_final_data_0715/local/mega_coffee_1
μš”μ¦˜ μΉ΄νŽ˜μ—μ„œ λ³΅μˆ­μ•„ μ•„μ΄μŠ€ν‹°λ₯Ό 자주 λ§ˆμ‹œλŠ”λ° 당이 많이 λ“€μ–΄ 살이 μ°”κΉŒλ΄ 걱정돼. 메가컀피 μ‚¬μ΄νŠΈμ—μ„œ λ³΅μˆ­μ•„ μ•„μ΄μŠ€ν‹°μ˜ λ‹Ήλ₯˜λ₯Ό ν™•μΈν•΄μ„œ μ•Œλ €μ€˜.
https://mega-mgccoffee.com/
easy
[ "information_lookup" ]
easy_final_data_0715/local/naver_cafe_1
μΉ΄νŽ˜μ—μ„œ μš”μ¦˜ μΈκΈ°μžˆλŠ” κ²Œμ‹œκΈ€λ“€μ΄ μ–΄λ–€ λ‚΄μš©μΈμ§€ κΆκΈˆν•΄. λ„€μ΄λ²„μΉ΄νŽ˜μ—μ„œ 인기글 탭에 λ“€μ–΄κ°„ λ‹€μŒ μ‹€μ‹œκ°„ 인기글 μ΅œμƒλ‹¨ κΈ€μ˜ 제λͺ©μ„ μ•Œλ €μ€˜.
https://section.cafe.naver.com/ca-fe/home
medium
[ "ranking", "site_interaction" ]

πŸ‡°πŸ‡· Ko-Online-Mind2Web

A live Korean web-agent benchmark, built manually by native Korean annotators.

πŸ“‹ 312 tasks Β· 🌐 104 websites Β· ✍️ ~480 annotator-hours Β· πŸ”΄ Live evaluation

πŸ“Œ Overview

Ko-Online-Mind2Web is a live web-agent benchmark for Korean websites. It contains 312 tasks across 104 website hostnames, written directly in Korean. The published task interface follows established international benchmark protocols, specifically Online Mind2Web: an agent starts from a given URL, operates the real website, and is judged on whether the final outcome satisfies the request.

✍️ 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
πŸ‘€ 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 live task-completion setup follows Online Mind2Web: a natural-language task and starting URL define the objective, and task success depends on the outcome achieved on a real website. The Korean requests and demonstrations were manually collected for this release. A task can be completed through multiple valid interaction paths.

How the benchmark was constructed and verified

  1. Direct Korean task creation. Korean annotators wrote realistic requests for Korean websites and Korean-language interfaces, covering constrained search, ranking, information retrieval, and multi-page interactions.
  2. Manual browser demonstrations. Korean human annotators personally performed and recorded all 312 released tasks on the real websites, producing human reference demonstrations for every task.
  3. Manual task eligibility review. Korean human annotators manually checked login requirements, inaccessible or blocked websites, and dated requests. Through this human review, the no-login pool of 326 tasks was reduced to 312 tasks by excluding 14 date-bound cases. The annotators also reviewed and revised the instruction wording where needed.
  4. Release verification. All 312 task IDs are unique and have nonempty start URLs, valid difficulty labels, and nonempty category assignments. The exported schema, example membership, and reported distributions are checked against the actual release file.

πŸ“Š Dataset statistics

Domain distribution

Domain Share
Ecommerce & shopping 26.6%
Science & education 13.1%
Computers & tech 9.0%
Arts & entertainment 7.4%
Food & drink 6.1%
Travel & tourism 6.1%
Law & government 5.8%
Reference materials 5.8%
News & media 4.2%
Other 16.0%

Difficulty distribution

Category Count Share
easy 97 31.09%
medium 120 38.46%
hard 95 30.45%

Difficulty labels (easy, medium, and hard) are assigned according to the length of the human reference trajectory for each task.

Task-category distribution

Stored label Meaning Tasks
filter Constraint-based search 104
ranking Rank-based selection 111
return_navigation Multi-page workflow 141
site_interaction Functional interaction 111
information_lookup Information retrieval only 90

Category labels can overlap: 176 tasks have multiple labels and the 312 tasks have 557 assignments in total. The current category allocation was curated post hoc after inspecting evaluation outcomes; it is an exploratory analysis taxonomy. statistics.json includes these distributions and per-hostname 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 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-Online-Mind2Web",
    repo_type="dataset",
    local_dir="Ko-Online-Mind2Web",
    allow_patterns=["README.md", "statistics.json", "online_mind2web_ko_tasks.json", "images/*"],
    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-Online-Mind2Web", split="test", token=True)

Evaluation

Run the agent from each start_url with the corresponding instruction and judge whether the final website outcome satisfies the request. Report success rate over all 312 tasks, the evaluation date, and the judging procedure.

Contact

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

Contributor

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{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}
}
@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}
}
Downloads last month
23

Collection including KORMo-VL/Ko-Online-Mind2Web

Paper for KORMo-VL/Ko-Online-Mind2Web