mobileforge-anonymous's picture
Publish sanitized anonymous files
8e3ac9f verified
|
Raw
History Blame Contribute Delete
4.41 kB
---
license: apache-2.0
language:
- en
pretty_name: MobileForge Generated Tasks
tags:
- mobileforge
- mobile-gui-agent
- android
- task-generation
- curriculum-learning
- mobilegym
task_categories:
- text-generation
size_categories:
- 1K<n<10K
viewer: false
configs:
- config_name: default
data_files:
- split: train
path: generated_tasks_26020301-all.csv
---
# MobileForge Generated Tasks
Anonymous project: https://mobileforge-anonymous.github.io/
Anonymous code: https://github.com/mobileforge-anonymous/MobileForge
This dataset contains the consolidated task pool generated by MobileGym-Curriculum from target-app exploration trajectories. These tasks are used by MobileForge for rollout collection and annotation-free adaptation.
<!-- mobileforge-release-inventory:start -->
Release inventory `payloads`: files=1; bytes=2028399; sha256=66d0c571583fad088279323c7332ef99ad1d21a3df76f0d969ce698c62f4ebfb
<!-- mobileforge-release-inventory:end -->
## Dataset summary
| File | Rows | Apps | Size | Description |
| --- | ---: | ---: | ---: | --- |
| `generated_tasks_26020301-all.csv` | 3,249 | 20 | 1.93 MB | Consolidated AndroidWorld-side MobileForge task pool. |
The task pool is generated from real target-app exploration traces. Each row describes an executable mobile GUI task candidate together with its source app, source trajectory id, task-generation metadata, and coarse feasibility signals from the task-generation stage.
## Load with `datasets`
```python
from datasets import load_dataset
ds = load_dataset("mobileforge-anonymous/mobileforge-generated-tasks", split="train")
print(ds[0])
```
Load with pandas:
```python
import pandas as pd
df = pd.read_csv("generated_tasks_26020301-all.csv")
print(df[["app_name", "task_description"]].head())
```
## Columns
| Column | Description |
| --- | --- |
| `task_identifier` | Public task id formed from the source trajectory id and task index. |
| `task_description` | Natural-language mobile GUI task. |
| `golden_steps` | Estimated number of reference steps from task generation. |
| `app_package` | Android package name. |
| `app_name` | Human-readable app name. |
| `trajectory_id` | Source exploration trajectory id. |
| `original_goal` | Exploration goal that grounded the generated task. |
| `task_reasonable` | Whether the task was judged reasonable by the generation pipeline. |
| `task_completed` | Whether the source trajectory completed the grounding goal. |
| `task_id` | Per-trajectory task index. |
| `difficulty_level` | Optional difficulty annotation; empty in this release. |
| `core_functionality` | Generated functionality label. |
| `variation_type` | Type of task variation, such as scenario application or step progression. |
| `prerequisites` | Preconditions assumed by the generated task. |
## App distribution
| App | Tasks |
| --- | ---: |
| Files | 258 |
| Pro Expense | 247 |
| Broccoli Recipe | 241 |
| Simple SMS Messenger | 240 |
| Markor | 233 |
| Clock | 215 |
| Retro Music | 189 |
| Contacts | 166 |
| VLC | 161 |
| Chrome | 157 |
| Tasks | 150 |
| Simple Draw Pro | 145 |
| OsmAnd | 136 |
| Joplin | 130 |
| Simple Calendar Pro | 107 |
| OpenTracks | 104 |
| Settings | 102 |
| Camera | 94 |
| Audio Recorder | 91 |
| Simple Gallery Pro | 83 |
## Generation metadata
- Reasonable tasks: 3,213 / 3,249.
- Tasks grounded in completed source trajectories: 1,691 / 3,249.
- Common variation types include `scenario_application`, `step_progression`, `multi-step_workflow`, and `parameter_change`.
The file `manifest.json` provides machine-readable file metadata, including the row count and column list.
## Relationship to other MobileForge artifacts
This dataset is the bridge between:
1. `mobileforge-anonymous/mobileforge-exploration-trajectories`: target-app exploration traces.
2. `mobileforge-anonymous/mobileforge-training-data`: hint-contextualized step-level GRPO samples produced after rollout and hierarchical evaluation.
3. `mobileforge-anonymous/mobileforge-benchmark-results`: AndroidWorld and MobileWorld evaluation artifacts.
## Limitations
The tasks are automatically generated from exploration trajectories and may include infeasible or environment-state-dependent assumptions. MobileForge handles this downstream through rollout feedback, task filtering, and step-level policy optimization.
## Citation
Citation information is withheld during double-blind review.