Hugging Face's logo

Datasets:
lamshiu
/
gui-world-model-data

Size:
100K<n<1M
Tags:
world-model
gui
computer-use
screenshots
games
License:
Dataset card Data Studio Files
xet
 Community
gui-world-model-data / README.md
lamshiu's picture
lamshiu
Add files using upload-large-folder tool
1feacd5 verified
preview code
|
Raw History Blame Contribute Delete
3.38 kB
metadata 
license: cc-by-4.0
pretty_name: GUI World Model Data
size_categories:
  - 100K<n<1M
tags:
  - world-model
  - gui
  - computer-use
  - screenshots
  - games

GUI World Model Data

Interaction data for training small action-conditioned world models of graphical user interfaces: a model sees a screen and an action (click, drag, double-click, key) and predicts the next screen. Every app and game here was driven automatically in headless Chromium (Playwright); each record is screen -> action -> next screen.

Raw data only: screenshots, app states, actions. No model features are included (they are recomputed from the screenshots with a frozen image encoder).

Contents

Folder What Size
level0/ Two-player tic-tac-toe, clicks 5,095 unique screens, 100,000 clicks
settings/ Small settings app: toggles, slider drag, dropdown menu, reset 524 screens, 30,000 actions
files/ Small file manager: open folders, drag files onto folders, double-click preview (has hidden state) 1,885 states, 40,000 actions
snake/ Snake on an 8x8 board, keys, random food 40,001 frames, 40,000 steps
snake_human/ One short human Snake session 272 frames
paint/ Paint app: drag strokes with colours 40,004 frames, 4 workers
tetris/ Tetris-style game (21 gravity speeds, 8 keys), random play 80,008 frames, 8 workers
tetris_expert/ Same game, played by a scripted expert 80,008 frames, 8 workers
tinyapps/ TinyApps: 1,100 procedurally generated small apps (toggles, checkboxes, counters, radio groups, dropdowns, dialogs, tabs; random layout, theme, labels, wiring). States and actions only - no screenshots 69,892 states, 330,000 clicks

Model checkpoints (*.pt, PyTorch) trained on these folders are included next to their data.

Formats

  • Screenshots are packed into tar shards img-000.tar, img-001.tar, ... (at most ~1 GB each) next to the folder's JSON. Inside: 0.png, 1.png, ... (Tetris: lossless WebP, 0.webp, ...; bit-identical to the original PNGs). File number = state / frame index. Untar into an img/ folder.
    • 224 x 224 RGB for all apps; Tetris frames are 620 x 620 crops.
  • Apps with unique states (level0, settings, files): states.json (one entry per unique state; screenshot i shows state i) and transitions (transitions.npz: src, x, y, dst for tic-tac-toe; transitions.json: [src, kind, x1, y1, x2, y2, dst] with kind 0 = drag, 1 = click, 2 = double-click, 3 = type).
  • Games (snake, paint, tetris*): frame sequences; states.json[t] describes frame t, actions.json[t] is the action taken after frame t; episodes.json gives the episode of each frame where present.
  • TinyApps (tinyapps/apps_w*.json): a list of apps {seed, spec, states, trans} with trans = [src, x, y, dst] (click position in pixels of the 224 x 224 page). Screens are re-rendered from the states by the generator page (tinyapp.html#SEED, window.__set(state)), which is deterministic.

Notes

  • files/ has hidden state: what is inside a closed folder is not visible, so many states share one screenshot.
  • Tetris frames carry faint renderer noise (low-amplitude pixel differences between otherwise equal frames).
  • Code (collectors, generator, models) will be released with the accompanying paper.