FruitPunchSamuraiG/openapps-buttons-v5-captions

OpenApps buttons-v5, deduplicated frames for captioning

Screenshots of the OpenApps web environment, collected for training a visual latent world model (JEPA + SIGReg) for web navigation, and exported here for captioning.

36,785 unique frames, deduplicated from 274,289 recorded frames (13.4%). Viewport is 1024x640, stored as JPEG.

Why the frames are deduplicated

Collection is matched-state: one exactly reproducible screen (app, theme, content, seed, scroll depth) is replayed once per available action, with a byte-identical padding prefix on every branch, so the branches differ only by which action fires. That makes each group a controlled experiment, and it also means the same screen is stored many times. Deduplication here is on the JPEG bytes, so it is exact rather than approximate.

refs.csv maps every original (episode_idx, step_idx) back to a frame, so a caption written once rejoins every occurrence in the training data.

Files

• data/train-*.parquet - 8 shards, the frames themselves with their metadata
• manifest.csv - the same metadata without the images, for quick filtering
• refs.csv - frame_id, lance, episode_idx, step_idx, action_name

from datasets import load_dataset
ds = load_dataset("FruitPunchSamuraiG/openapps-buttons-v5-captions", split="train")
ds[0]["image"]        # PIL.Image, 1024x640
ds[0]["action_names"] # list[str]

Columns

The parquet files carry an image column plus every manifest.csv column except path.

column	meaning
frame_id	first 16 hex of the SHA1 of the JPEG bytes
app	which screen (see table below)
theme	appearance variant
view	recorder's view label
url	path within the app (host and ephemeral port stripped)
n_refs	how many recorded frames this one screen covers
action_names	actions whose episodes contain this frame, |-separated
delta_pct	% of pixels the action changed, measured at collection time

Telling "before" from "after"

action_names is the useful column. A frame from a group's shared prefix lists all of that group's actions, because every branch replayed it. A frame that is one action's outcome lists exactly one. So filtering to single-action rows gives the "after" frames, and their group's multi-action rows give the matching "before" - no index arithmetic needed.

Coverage

app	frames	themes
buttons_calendar	6,903	black_and_white, challenging_font, dark_theme, default
buttons_codeeditor	2,683	black_and_white, challenging_font, colorblind_access, dark_theme, default
buttons_home	7,377	black_and_white, challenging_font, dark_theme, default
buttons_map	11,619	black_and_white, challenging_font, dark_theme, default
buttons_messages	2,973	black_and_white, challenging_font, dark_theme, default
buttons_todo	5,230	black_and_white, challenging_font, dark_theme, default

Actions present: Agenda, NewFile, NewFolder, Return, ZoomIn, ZoomOut, event, icon_calendar, icon_codeeditor, icon_maps, icon_messages, icon_todo, null, scroll_dn, scroll_up, thread

Caveats

• delta_pct is measured against the frame the action was applied to, not against this frame, so it is a property of the action, not of the image.
• map frames pull tiles from an external OSM server, so some carry blank or partially loaded tiles.
• Content within an app is randomized per state (todo list length, calendar events, message threads, editor file lists), which is deliberate: it is the same-task / different-content axis the world model is evaluated on.