KO-Web-Understanding

Korean web screen understanding data used to train the KORMo-Web agents. Each subset is one task; pick it by name.

Text only. Screenshots are not included; image_files gives the file name of the screen(s) each row was made from.

from datasets import load_dataset
ds = load_dataset("KORMo-VL/ko-web-understanding", "grounding", split="train")

Subset	Rows	Task
page_caption	1,419,854	Grounded description of the whole screen (element names with 0–1000 coordinates)
screen_qa	506,157	Question answering about what is on the screen
ocr_read_text	505,638	Read the text in a given region or the whole screen
grounding	1,517,176	text_to_point / point_to_text / appearance_to_point (0–1000 coordinates)
screen_change_caption	498,731	Describe what changed between two screens and how
action_between_screens	499,078	Which action (click / scroll) turns screen 1 into screen 2
action_between_screens_think	498,517	Same as above, with reasoning
next_action	505,550	Next action toward a goal (click)
next_screen_prediction	506,005	Predict the screen after a click
next_screen_prediction_detailed	505,692	Predict the screen after a click or scroll, in detail
Total	6,962,398

Fields

Field	Description
id	Unique row id
lang	ko
task / subtask	Task name and variant (e.g. click, scroll, text_to_point)
question	User prompt
answer	Target answer. For action tasks, a JSON tool call such as {"name": "click", "arguments": {"x": 500, "y": 300}}
reasoning	Reasoning text when the task has one, else null
image_files	Screenshot file name(s) the row refers to

Coordinates are normalized to 0–1000.

Notes

• Screens come from public Korean websites. Screens captured while logged in (my page, cart, checkout, mail, social media, community sites, etc.) were removed.
• Session, token and password values inside URLs were replaced with [MASKED].
• page_caption keeps all mid-length captions; other subsets are random samples of about 500k rows each (grounding: about 500k per variant).