cua-lite/MobileGym

MobileGym train-split model rollouts; every non-OOB trajectory kept, quality gates tagged in metadata.others.exclude_reason (strict SFT filter: not exclude_reason and episode_return>=1.0).

Origin

Load via datasets

from datasets import load_dataset

# entire dataset
ds = load_dataset("cua-lite/MobileGym")

# just one named subset (config)
ds = load_dataset("cua-lite/MobileGym", "GELab-Zero-4B-preview")

You can also filter by metadata.platform / metadata.task_type / metadata.others.* after loading; every row carries a rich metadata struct (see schema below).

Schema

Each row has these columns:

column	type	notes
images	list[Image]	embedded PNG/JPEG bytes; HF viewer renders thumbnails
messages	list[struct]	OpenAI-style turns with role + structured content
metadata	struct	{platform, task_type, extra_tool_schemas, valid_actions, others{...}}

Coordinate values in messages are normalized to [0, 1000] integers.

Image-dedup (grounding.* / understanding cohorts). These cohorts are single-image-per-row and many rows share the same screenshot, so to avoid re-embedding identical image bytes once per instruction they are stored folded: one row per unique screenshot (image embedded once), carrying an extra _folded column — a JSON string with the authoritative list of {messages, metadata} members for that screenshot. The row's top-level messages is the members concatenated for viewer convenience. use cohorts are not folded. Use lite.data.hf.download to consume this repo — it unfolds automatically back to one row per instruction; reading the parquet directly yields the folded form.

Layout

<platform>/<task_type>/<split>/shard-NNNNN-of-NNNNN.parquet                  # single-variant cohort
<platform>/<task_type>/<split>/<variant>/shard-NNNNN-of-NNNNN.parquet        # multi-variant cohort

• platform ∈ {desktop, mobile, web}
• task_type ∈ {understanding, grounding.action, grounding.point, grounding.bbox, use} — used verbatim as the dir component
• HF config names are <platform>.<task_type> by default (e.g. mobile.grounding.action) — UNLESS the dataset was staged with --config-names, which sets verbatim, explicitly-chosen config names (see the configs: block above for the authoritative list). The agent registry lookup key in code is <agent>@<platform>@<task_type> (e.g. qwen3_vl@[email protected]); only this user-facing token uses . between platform and task_type, because @ triggers a 403 on the dataset-viewer's signed image URLs.
• HF split names stay train / validation (the datasets library blacklists <>:/\|?* in split names; everything else is fine in config_name)
• validation is an in-distribution held-out slice (never used in training); test is reserved for out-of-distribution benchmark datasets

Stats

platform	task_type	variant	train	validation
mobile	use	GELab-Zero-4B-preview	160	0
mobile	use	MAI-UI-2B	160	0
mobile	use	MAI-UI-8B	160	0
mobile	use	Qwen3-VL-2B-Instruct	160	0
mobile	use	Qwen3-VL-32B-Instruct	160	0
mobile	use	Qwen3-VL-4B-Instruct	160	0
mobile	use	Qwen3-VL-8B-Instruct	160	0
mobile	use	Qwen3.5-27B	160	0
mobile	use	Qwen3.5-2B	160	0
mobile	use	Qwen3.5-4B	160	0
mobile	use	Qwen3.5-9B	160	0
mobile	use	UI-TARS-1.5-7B	160	0
mobile	use	UI-TARS-7B-DPO	160	0
mobile	use	UI-Voyager	160	0
mobile	use	gpt-5.4	319	0
mobile	use	gpt-5.5	320	0

Local mirror & SFT export

For local workflows (SFT export, dedup, mixing across datasets), use lite.data.hf.download to mirror this repo back to the canonical local layout:

$CUA_LITE_DATASETS_ROOT/cua-lite/MobileGym/
  images/<hash[:2]>/<hash>.<ext>                          # content-addressed image store
  <platform>/<task_type>/<split>[/<variant>].parquet      # rows reference images by relative path

Rows in the local parquet have images: list[str]; bytes are extracted to the image store. lite.train.export.export_sft consumes the local form directly with --image-root=$CUA_LITE_DATASETS_ROOT.

• Total unique images: 20,677
• Image store size: 10.48 GB

Notes

Staged via lite.data.hf.stage from rollout log-roots: .data/rollout/mobilegym/gpt-5.4/398f9e5ab/train_annotated, .data/rollout/mobilegym/gpt-5.5/398f9e5ab/train_annotated, .data/rollout/mobilegym/mai_ui_2b/8227e4d4e/train_annotated, .data/rollout/mobilegym/mai_ui_8b/8227e4d4e/train_annotated, .data/rollout/mobilegym/qwen3_5_27b/8227e4d4e/train_annotated, .data/rollout/mobilegym/qwen3_5_2b/8227e4d4e/train_annotated, .data/rollout/mobilegym/qwen3_5_4b/8227e4d4e/train_annotated, .data/rollout/mobilegym/qwen3_5_9b/398f9e5ab/train_annotated, .data/rollout/mobilegym/qwen3_vl_2b/398f9e5ab/train_annotated, .data/rollout/mobilegym/qwen3_vl_32b/398f9e5ab/train_annotated, .data/rollout/mobilegym/qwen3_vl_4b/398f9e5ab/train_annotated, .data/rollout/mobilegym/qwen3_vl_8b/398f9e5ab/train_annotated, .data/rollout/mobilegym/step_gui/8227e4d4e/train_annotated, .data/rollout/mobilegym/ui_tars/8227e4d4e/train_annotated, .data/rollout/mobilegym/ui_tars_15/8227e4d4e/train_annotated, .data/rollout/mobilegym/ui_voyager/8227e4d4e/train_annotated (row filter: lambda m: 'oob_coordinate' not in (m.others.get('exclude_reason') or '')).

License & citation

other