KMK040412/aitw-processed-labeled-full

AiTW Processed Full with App Labels

This repository contains a full processed Android in the Wild (AiTW) mirror together with an app-labeled step index, official split assignment by episode_id, major-app statistics, and a ready-to-train Gmail subset.

Why This Exists

AiTW is large and not easy to navigate by app. The original labels contain useful fields such as goal_info, current_activity, and action coordinates, but users often need extra processing before they can answer questions like:

• Which apps dominate the dataset?
• Which episode ids belong to the official train/validation/test split?
• Where are all Gmail examples?
• Which package/activity does each step come from?
• How can app-specific SFT subsets be built without scanning the whole archive again?

This dataset answers those questions by publishing the processed source archive and derived labels together.

Source And Compatibility

Source candidate used for the processed archive:

jacklishufan/aitw

Important compatibility details:

• The source mirror is a split POSIX tar archive, not the official raw GCS TFRecord format.
• Internal layout is processed AiTW:

processed/<domain>/episode_<episode_id>/step_<step_id>_raw.jpg
processed/<domain>/episode_<episode_id>/step_<step_id>_anno.jpg
processed/<domain>/episode_<episode_id>/step_<step_id>_payload.txt

• Each step_*_payload.txt preserves official-compatible labels:

episode_id, step_id, goal_info, current_activity, action_text,
ex_action_type, touch_x, touch_y, lift_x, lift_y, type_text,
image_height, image_width, device_type

• Official Google Research AiTW split JSON files were downloaded from:

gs://gresearch/android-in-the-wild/splits/

• The published split column is assigned by exact episode_id matching against the official standard.json split.

Repository Layout

full_processed_tar/
  a.tar.part-000 ... a.tar.part-033

viewer_examples/
  gmail_mobile_use_examples.parquet
  app_index_sample.parquet

app_index/
  aitw_app_index.parquet
  aitw_app_index_stats.json
  major_apps.csv
  major_packages.csv

gmail/
  aitw_gmail_google_apps_train.parquet
  aitw_gmail_google_apps_validation.parquet
  aitw_gmail_google_apps_test.parquet

manifests/
  gmail_payloads.stats.json
  gmail_payloads_episodes.jsonl

official_splits/
  standard.json
  unseen_android_version.json
  unseen_domain.json
  unseen_subject.json
  unseen_verb.json

scripts/
  build_app_index.py
  scan_gmail_payloads.py
  extract_gmail_episodes.py
  convert_gmail_to_parquet.py
  create_viewer_examples.py

README_COMPATIBILITY.md
PROCESSING_REPORT.md
README.md

Dataset Viewer

The dataset card defines explicit Hugging Face Dataset Viewer configs:

• 000_examples: small Gmail examples with screenshot image, instruction, action label, and mobile_use JSON target. This config is intentionally named with a numeric prefix so it appears first in the Dataset Viewer.
• 001_app_index_sample: small app-index examples for browsing the app/package labeling schema.
• gmail_sft: full Gmail train/validation/test parquet subset.
• app_index: full step-level app/package index.

The large split tar archive is intentionally not included in the viewer configs. It remains available under full_processed_tar/ for full reconstruction and streaming extraction, but the viewer opens the small examples first.

Example row fields in 000_examples:

id
split
episode_id
step_id
instruction
current_activity
image
target_json
action
coordinate
coordinate2
action_text
image_width
image_height

Example strict tool-call target:

{"name":"mobile_use","arguments":{"action":"click","coordinate":[695,793]}}

Full Processed Archive

The full processed archive is included under full_processed_tar/ as the original split tar parts.

To list files:

cat full_processed_tar/a.tar.part-* | tar -tf - | head

To extract everything:

mkdir -p extracted
cat full_processed_tar/a.tar.part-* | tar -xf - -C extracted

To extract a single indexed member:

MEMBER="processed/google_apps/episode_.../step_0_raw.jpg"
cat full_processed_tar/a.tar.part-* | tar -xOf - "$MEMBER" > step_0_raw.jpg

Full App Index

app_index/aitw_app_index.parquet contains one row per AiTW step. It does not duplicate image bytes. Instead it links each row back to the archive member through tar_member.

Columns:

domain
split
episode_id
step_id
package
activity
app_label
label_source
goal_info
action_text
action_family
ex_action_type
touch_x
touch_y
lift_x
lift_y
type_text_present
image_height
image_width
tar_member

Total indexed steps: 5,687,349

Domain distribution:

{
  "general": 85413,
  "google_apps": 4903483,
  "install": 250058,
  "web_shopping": 365253,
  "single": 83142
}

Official standard split distribution:

{
  "train": 4485992,
  "validation": 560607,
  "test": 557608,
  "unknown": 83142
}

Major Apps

The app label is derived from the current activity package first. If a package is unknown or not mapped, goal keywords are used as a secondary fallback. Long-tail unknown packages fall back to a readable package tail.

Important disambiguation:

• Gmail is com.google.android.gm.
• Google Play Services is com.google.android.gms.
• These two are explicitly separated to avoid false Gmail matches.

Top apps by step share:

Rank	App label	Steps	Episodes	Step share
1	Chrome	1,709,697	208,667	30.06%
2	Android Settings	1,319,064	169,766	23.19%
3	Gmail	534,517	61,302	9.40%
4	Google Calendar	318,490	39,606	5.60%
5	Google Maps	300,599	45,921	5.29%
6	Clock	258,059	32,841	4.54%
7	Google Play Store	209,388	22,837	3.68%
8	Google Play Services	203,982	17,612	3.59%
9	Google Photos	203,978	23,902	3.59%
10	Google Search	196,580	33,810	3.46%
11	YouTube	83,587	10,068	1.47%
12	Contacts	70,688	8,702	1.24%
13	Nexuslauncher	51,237	14,320	0.90%
14	Latin	36,728	2,487	0.65%
15	Permissioncontroller	22,725	2,531	0.40%

The complete ranking is available in:

app_index/major_apps.csv
app_index/major_packages.csv

Gmail Ready-To-Train Subset

Gmail was extracted as an image-included parquet subset because it is a high-value app for GUI-agent overfitting/SFT.

Gmail extraction rule:

• domain: google_apps
• exact Gmail package match: com.google.android.gm
• or goal/activity contains the word gmail
• excludes com.google.android.gms Google Play Services false positives

Gmail statistics:

Gmail steps: 607,799
Gmail episodes: 68,657
train steps: 486,296
validation steps: 61,032
test steps: 60,471

Gmail parquet columns:

id
source
domain
episode_id
step_id
split
instruction
current_activity
image
image_path
anno_image_path
target_json
action_text
ex_action_type
touch_x
touch_y
lift_x
lift_y
type_text
image_height
image_width

target_json is converted into strict mobile_use JSON format, for example:

{"name":"mobile_use","arguments":{"action":"click","coordinate":[695,793]}}

or:

{"name":"mobile_use","arguments":{"action":"swipe","coordinate":[658,888],"coordinate2":[745,361]}}

Processing Pipeline

The data was processed on a Vultr bare-metal server with:

CPU: AMD EPYC 4345P, 8 cores / 16 threads
RAM: 64GB
Storage: NVMe, /data working directory

Processing steps:

1. Downloaded the split tar archive from the public HF mirror.
2. Downloaded official AiTW split JSON files from Google Research GCS.
3. Stream-scanned the full tar archive without extracting all files.
4. Parsed every step_*_payload.txt JSON file.
5. Extracted package name from current_activity.
6. Assigned readable app_label from package map and goal fallback.
7. Assigned official split using episode_id from standard.json.
8. Wrote the full step-level app index as parquet.
9. Created major_apps.csv and major_packages.csv summaries.
10. Extracted Gmail episodes only.
11. Converted Gmail steps to image-included train/validation/test parquet files.
12. Uploaded the full processed tar archive and labels together to this public repository.

How To Use

Read the global app index:

import pandas as pd
idx = pd.read_parquet("app_index/aitw_app_index.parquet")
print(idx[["app_label", "package", "split", "goal_info"]].head())

Find all Gmail rows in the full index:

gmail = idx[idx.app_label == "Gmail"]

Load the Gmail SFT subset:

import pandas as pd
train = pd.read_parquet("gmail/aitw_gmail_google_apps_train.parquet")
print(train.iloc[0]["instruction"])
print(train.iloc[0]["target_json"])

Caveats

• This is a processed mirror-derived dataset, not the official raw TFRecord release.
• The full archive is large: about 331GiB in split tar parts.
• app_label is a practical normalization, not an official AiTW field.
• Some long-tail install apps still use package-tail labels and may benefit from manual cleanup.
• Gmail and Google Play Services are explicitly separated, but other Google settings flows may cross app boundaries during an episode.

Recommended Attribution

Please cite the original Android in the Wild dataset and the source mirror used for the processed archive. This repository is a derived labeling/indexing and app-specific packaging pass.