Release inventory raw: files=539; bytes=3995527793; sha256=798b58729f5bfee7a0f929371500d3532f15343524a5e14bb14fc4132480d963 Release inventory final: files=6071; bytes=2474994885; sha256=3a336c6dc6724e2561d24165ba714eab3ea8f00e9ee10fc7037fce0eadf2e415

Role in MobileForge

MobileForge starts from annotation-free target-app exploration. The exploration stage collects function-aware interaction trajectories from 20 Android apps. MobileGym-Curriculum then mines these traces into executable task candidates, which are later used for rollout, hierarchical evaluation, and HiFPO training.

The Hugging Face Dataset Viewer is disabled because the repository contains compressed trajectory objects and APK files. The lightweight JSON summaries in final/ are the recommended entry point.

Splits

Split	Path	Apps	Files	Size	Description
Sanitized exploration artifacts	`raw/`	20	539	3.72 GB	Sanitized and deterministically rebuilt exploration artifacts, including compressed trajectory objects and app metadata.
Parsed summaries	`final/`	20	6071	2.31 GB	Parsed/final exploration outputs used for curriculum generation.

High-level statistics

The parsed final/ summaries contain 520 exploration trajectories over 20 apps, with 511 successful trajectories and 5,839 total interaction steps.

App	Package	Trajectories	Successful	Steps
Retro Music	`code.name.monkey.retromusic`	38	35	452
Camera	`com.android.camera2`	17	15	111
Chrome	`com.android.chrome`	26	26	359
Settings	`com.android.settings`	15	15	131
Pro Expense	`com.arduia.expense`	25	24	200
Audio Recorder	`com.dimowner.audiorecorder`	16	16	232
Broccoli	`com.flauschcode.broccoli`	63	63	594
Contacts	`com.google.android.contacts`	30	30	252
Clock	`com.google.android.deskclock`	33	33	308
Files	`com.google.android.documentsui`	41	41	446
Calendar	`com.simplemobiletools.calendar.pro`	15	15	212
Draw	`com.simplemobiletools.draw.pro`	27	25	242
Gallery	`com.simplemobiletools.gallery.pro`	14	13	196
SMS Messenger	`com.simplemobiletools.smsmessenger`	28	28	288
OpenTracks	`de.dennisguse.opentracks`	16	16	221
Joplin	`net.cozic.joplin`	20	20	172
Markor	`net.gsantner.markor`	35	35	594
OsmAnd	`net.osmand`	23	23	368
Tasks	`org.tasks`	21	21	303
VLC	`org.videolan.vlc`	17	17	158

Directory layout

exploration-trajectories/
  manifest.json
  raw/
    <app_package>/
      app_info.json
      <trajectory_id>.pkl.zst
      <app_package>.apk
  final/
    <app_package>/
      app_info.json
      statistics.json
      trajectories_summary.json

app_info.json records the app name, package, version, and main activity. statistics.json summarizes app-level trajectory counts and step counts. trajectories_summary.json contains one record per explored trajectory with the trajectory id, depth, exploration goal, step count, success flag, raw file path, and start time.

Example trajectories_summary.json record:

{
  "trajectory_id": "260114181539.908106_D7E48636EB451E79",
  "depth": 1,
  "goal": "Allow the app to send notifications, then find and open the contact ...",
  "steps_count": 12,
  "success": true,
  "file_path": "exploration_output_parallel_26011201/com.google.android.contacts/260114181539.908106_D7E48636EB451E79.pkl.zst",
  "start_time": "2026-01-14T18:15:39"
}

Download examples

Download only parsed summaries:

from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="mobileforge-anonymous/mobileforge-exploration-trajectories",
    repo_type="dataset",
    allow_patterns=[
        "manifest.json",
        "final/*/app_info.json",
        "final/*/statistics.json",
        "final/*/trajectories_summary.json",
    ],
)

Load a summary file:

import json
from pathlib import Path

summary_path = Path(local_dir) / "final/com.google.android.contacts/trajectories_summary.json"
records = json.loads(summary_path.read_text())
print(records[0])

Notes

The raw/ split contains anonymized compressed exploration objects and the original third-party APKs for reproducibility.
The final/ split is usually sufficient for inspecting exploration coverage and reproducing MobileGym-Curriculum task mining.
Compressed pickle objects should be treated as executable Python serialization. Load them only in a trusted research environment.
APK redistribution may be subject to the upstream app licenses. Use this dataset for research and reproducibility.

MobileForge Exploration Trajectories