Dataset Viewer
Auto-converted to Parquet Duplicate
The dataset viewer is not available for this split.
Parquet error: Scan size limit exceeded: attempted to read 455854069 bytes, limit is 300000000 bytes Make sure that 1. the Parquet files contain a page index to enable random access without loading entire row groups2. otherwise use smaller row-group sizes when serializing the Parquet files
Error code:   TooBigContentError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

GUI-Owl AndroidWorld SFT Mix — FULL / generalization

Purpose: AndroidWorld (116-task) SFT for a GUI-Owl-1.5-2B block-diffusion VLA. This dataset is an action-balanced, source-mixed SFT corpus assembled from five Android GUI-agent trajectory sources. It is built for in-domain supervised fine-tuning ahead of RL. The mix deliberately includes AndroidWorld task-family coverage (via the openmobile anchor, whose app field holds AW task-family names) and accepts in-domain overlap by design.

This version: FULL

Broadest mix (~773k steps). Maximizes diversity and visual ballast: AITW ~26%, gui_odyssey ~28%, openmobile ~22%. Lower scarce-verb floors (open 2.5%, answer 0.6%). Use when you want the model to generalize across many Android apps before specializing.

  • Total steps: 773,359
  • Shards: 194 parquet files (balanced-*.parquet, ~4000 rows each)
  • Build seed: 7

Action distribution

action rows share
click 382,500 49.46%
type 127,500 16.49%
system_button 85,000 10.99%
swipe 68,000 8.79%
terminate 42,500 5.5%
wait 33,009 4.27%
open 21,250 2.75%
long_press 8,500 1.1%
answer 5,100 0.66%

Source mix

source rows share
gui_odyssey 218,159 28.21%
aitw 200,220 25.89%
openmobile 169,383 21.9%
androidcontrol_par 129,309 16.72%
amex 56,288 7.28%

Swipe direction (overall)

dir rows share
up 42,653 62.73%
down 15,291 22.49%
left 5,226 7.69%
right 4,830 7.1%

Note: a direction-balanced swipe subset (up40/down30/left17/right13) is enforced up to 19,069 swipes; the remainder keeps the natural (up-heavy) distribution as ballast, so the overall swipe split above is up-heavy by design.

Schema (unified 13-column)

source           string   # origin dataset: androidcontrol_par | amex | openmobile | gui_odyssey | aitw
src_key          string   # original sample key in the source corpus
app              string   # app / AndroidWorld task-family label (may be empty)
instruction      string   # natural-language goal for the episode
history          string   # prior-step context
screenshot       binary   # PNG/JPEG screen bytes
screen_width     int
screen_height    int
episode_id       string
target_json      string   # ground-truth action as JSON (canonical label)
action_type      string   # parsed action: click|type|swipe|system_button|terminate|wait|open|long_press|answer
coordinate       list     # parsed primary coordinate (normalized 0-1000) or null
coordinate2      list     # parsed secondary coordinate (swipe end) or null

Coordinates are normalized to a 0-1000 grid (origin top-left).

How the mix was built

Balanced via a parallel 2-stage builder (build_balanced_mix.py): actions and coordinates are parsed from target_json, then per-action keep_prob, per-(action, coordinate-bin) caps (max_frac_per_coordbin = 0.06, 8x8 grid), and per-(action, app) caps (max_frac_per_app = 0.22) are applied. Scarce AndroidWorld verbs (open / answer / long_press / terminate) are up-weighted via sampling-with-replacement (replication caps: {'open': 6, 'answer': 8, 'long_press': 8, 'terminate': 5}). Full provenance and exact targets are in balance_report.json.

Sibling datasets

Downloads last month
605