Datasetpublic

KMK040412/guiowl-curated-corpus

GUI-Owl Curated Corpus This dataset publishes the full curated mobile GUI-agent supervised fine-tuning corpus in a unified norm1000 mobile_use action format. Each row pairs a mobile UI screenshot with an instruction and a normalized target tool call for training GUI agents. The published files are the curated parquet shards as produced by the source canonicalizers. No parquet shards are merged, re-sharded, or sampled during upload. Sources Source Episodes… See the full description on the dataset page: https://huggingface.co/datasets/KMK040412/guiowl-curated-corpus.

sourceHugging Faceupdated 4mo agoView on Hugging Face

0likes334downloads

Dataset Card

GUI-Owl Curated Corpus

This dataset publishes the full curated mobile GUI-agent supervised fine-tuning corpus in a unified norm1000 mobile_use action format. Each row pairs a mobile UI screenshot with an instruction and a normalized target tool call for training GUI agents.

The published files are the curated parquet shards as produced by the source canonicalizers. No parquet shards are merged, re-sharded, or sampled during upload.

Sources

Source	Episodes	Steps	Step Share	Parquet Shards	Coord Out Of Range
`androidcontrol_par` (AndroidControl)	12,232	67,315	1.134%	36	0
`amex` (AMEX)	3,046	38,489	0.649%	24	0
`openmobile` (OpenMobile)	2,754	27,359	0.461%	24	4
`gui_odyssey` (GUI-Odyssey)	7,735	113,676	1.916%	24	4,866
`aitw` (AITW)	714,327	5,687,349	95.840%	504	0
Total	740,094	5,934,188	100.000%	612

Grand total: 740,094 episodes and 5,934,188 steps across 612 parquet shards.

Format

Parquet shards are stored at:

—data/<source>/<shard>.parquet
—reports/<source>_report.json
—manifest.json

All sources use the same mobile GUI-agent SFT schema. Important columns are:

Column	Type	Description
`source`	`string`	Canonical source identifier.
`episode_id`	`string`	Episode or trajectory identifier within the source.
`step_id`	`int`	Step index within the episode.
`image`	`binary`	JPEG screenshot bytes for the current mobile UI state.
`instruction`	`string`	Step-level instruction or task prompt.
`target_json`	`string`	Tool-call target using the mobile_use action format.
`action_type`	`string`	Normalized action label such as click, swipe, type, wait, or terminate.
`coordinate`	`list[int]`	Primary [x, y] coordinate in norm1000 space, nullable.
`coordinate2`	`list[int]`	Secondary [x, y] coordinate for swipes or drags, nullable.
`coord_mode`	`string`	Coordinate mode, expected to be guiowlnorm1000xy.
`task`	`string`	High-level task text when available.
`app`	`string`	Application name or package when available.
`history`	`string`	Serialized prior instruction/action context when available.
`screen_width`	`int`	Original screenshot width when available.
`screen_height`	`int`	Original screenshot height when available.
`quality_tags`	`list[string]`	Optional curation tags.

target_json contains a <tool_call> payload whose name is mobile_use. Coordinate actions use [x, y] points in guiowl_norm1000_xy, meaning both axes are normalized into the inclusive 0-1000 range.

Mixing

This repository publishes the full curated data. Source mixing ratios are applied at training time by the dataloader or sampler, not by sampling in this upload.

Reports

Each source report is copied from the corresponding curation output directory and includes the original episode count, step count, shard count, action distribution, coordinate statistics, and coordinate out-of-range count.