Webintosh

1,110,686 instruction ↔ bounding-box pairs over 14,763 annotated screenshots from 1,100 scripted task trajectories in a macOS-style web desktop.

Environment: https://github.com/ChenghengLi/webintosh

GUI grounding — mapping a natural-language reference onto the exact pixels of the control it names — is the bottleneck capability for computer-use agents. Webintosh provides that supervision at scale, densely: rather than one labeled target per screenshot, every visible control on every screen is labeled, averaging ~75 grounded elements per frame.

What is in it

The environment is a faithful macOS desktop reimplemented for the browser: a menu bar, a Dock, draggable and resizable windows with traffic-light controls, Launchpad, Spotlight, Control Center, and a suite of working applications — Finder, Safari, Google Chrome, Mail, Messages, Notes, Reminders, Calendar, Photos, Music, Weather, App Store, Terminal, Calculator, Visual Studio Code, System Settings, Stocks, Activity Monitor, Dictionary, Stickies, TextEdit, Clock and more.

Trajectories are scripted multi-step tasks that exercise those applications the way a person would:

• Open Stocks, inspect Alphabet Class A shares, minimize the window and restore it from the Dock
• Open Google Chrome, search Nikola Tesla from the omnibox, open the article, follow an internal link
• Change appearance between Dark and Light mode and cycle accent colors in System Settings
• Open Safari from the Dock, search Wikipedia for Golden Gate Bridge, open the first result
• Send a quick question to Leo Park and one to Mom in Messages

Each trajectory runs on one of seven Apple display profiles, so the corpus spans a wide range of resolutions and pixel densities rather than a single canonical screen:

device	viewport (CSS px)	trajectories
MacBook Air 15"	1440 × 932	213
MacBook Pro 16"	1728 × 1117	207
MacBook Air/Pro 13"	1280 × 832	203
MacBook Pro 14"	1512 × 982	196
Studio Display 5K	2560 × 1440	117
Pro Display XDR 6K	3008 × 1692	85
iMac 24" 4.5K	2240 × 1260	79

All screenshots are captured at native Retina density (2×), up to 6016 × 3384 px.

Three phrasing styles

Every element is labeled in exactly one of three registers, balanced at a third each. This lets a model learn that the same box can be referred to in structurally different ways, and supports negative-instruction training:

style	share	example
imperative	33.0%	"open the Wi-Fi pane", "create a new note"
description	33.4%	"the Calendar icon", "the Groceries sidebar row"
negative	33.6%	"do not select the Terminal icon beside Calculator"

Row format

{
  "id": "wtraj_0001__step_20__e17",
  "img_filename": "screenshots/wtraj_0001/step_20.png",
  "instruction": "select the Groceries list in the sidebar",
  "style": "imperative",
  "bbox": [x1, y1, x2, y2],
  "img_size": [1512, 982],
  "device_scale_factor": 2,
  "device": "MacBook Pro 14\"",
  "trajectory_id": "wtraj_0001",
  "step_index": 20,
  "task": "Open Finder and navigate to Documents",
  "role": "button",
  "dom_name": "Groceries",
  "is_action_target": true
}

bbox and img_size are in CSS pixels; the PNG is device_scale_factor × larger in each dimension. To draw on the image, multiply by device_scale_factor.

task gives the trajectory-level goal, so a row carries both local and global intent. is_action_target marks the element the trajectory actually clicked at that step (13,263 rows) — the subset usable for action prediction and step-level evaluation.

Element roles are dominated by real controls: 75.3% button, 11.5% span, 6.1% div, 4.4% a, 1.4% input, with the remainder switches, selects and text areas.

Splits

split	rows
train	995,036
validation	57,522
test	58,128

Split by trajectory_id — no trajectory contributes to more than one split, so a model cannot memorise a screen in training and be scored on it at test time.

Loading

from datasets import load_dataset
ds = load_dataset("Chengheng/Webintosh")
row = ds["train"][0]

Fetching an image and drawing its box:

from huggingface_hub import hf_hub_download
from PIL import Image, ImageDraw

path = hf_hub_download("Chengheng/Webintosh", row["img_filename"], repo_type="dataset")
im = Image.open(path)
d = row["device_scale_factor"]
x1, y1, x2, y2 = [v * d for v in row["bbox"]]
ImageDraw.Draw(im).rectangle([x1, y1, x2, y2], outline="red", width=6)

screenshots/ also contains 2,097 keyboard-step frames (press / type actions). They carry full trajectory metadata but have no click target, so they produce no rows in the splits — useful for action prediction and sequence modelling.

How it was built

Elements are extracted from the live DOM with their roles, accessible names and exact geometry, then labeled by a vision-language model (gpt-5.6-luna) under Set-of-Mark prompting. Each request pairs a zoomed crop of the region being labeled — with numbered boxes drawn on the elements in question — against a whole-screen thumbnail that supplies app and overlay context. Replies are decoder-constrained to a JSON schema.

Element lists are cleaned before any labeling. SVG internals, boxes under 100 px², and passive span/img children swallowed by an interactive ancestor are collapsed, so each visual control contributes one box rather than a stack of overlapping ones. Elements whose pixels are blank are dropped. Where an instruction would fit two different boxes on the same screen, both copies are removed rather than arbitrarily choosing one, and boxes less than 95% visible are excluded — a target that runs off-screen cannot be grounded against the image.

A visual cache keyed on an element's own pixels plus its surroundings reuses labels across the many recurrences of the same control, keyed on evidence rather than screen position.

Quality

Measured across the whole corpus, not a sample:

check	result
Name agreement (622k labels carrying an accessible name)	86.0%
Action target present	14,763 / 14,763 (100%)
Style balance	33.0 / 33.4 / 33.6
Style compliance	98.7%
SVG / sub-100px DOM noise	0%
Boxes outside their image	0
Trajectory leakage across splits	0
Duplicate row ids	0
Empty instructions	0

Name agreement is a floor, not a ceiling: it asks whether a label mentions the element's own DOM name, so a better label counts as a miss — MON27 → "the Calendar icon" is correct but scores as a disagreement. Visual verification against pixels ran 10/10 and 6/6 on anonymous elements, the hard case where the model has only the image to work from. Realistic label accuracy is around 90%.

Instructions average 5.2 words (median 5, p95 9), matching the register of ScreenSpot-style referring expressions.

Citation

@misc{webintosh2026,
  title  = {Webintosh: A GUI Grounding Dataset for Web Desktop Environments},
  author = {Chengheng Li-Chen},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Chengheng/Webintosh}
}

Environment source: github.com/ChenghengLi/webintosh