README.md

---
license: apache-2.0
language:
- en
pretty_name: MobileForge Benchmark Results
tags:
- mobileforge
- mobile-gui-agent
- androidworld
- mobileworld
- gui-agent
- benchmark-results
task_categories:
- reinforcement-learning
- text-generation
viewer: false
---

# MobileForge Benchmark Results

Anonymous project: https://mobileforge-anonymous.github.io/  
Anonymous code: https://github.com/mobileforge-anonymous/MobileForge

This dataset contains the evaluation artifacts used by **MobileForge: Annotation-Free Adaptation for Mobile GUI Agents with Hierarchical Feedback-Guided Policy Optimization**.

It includes AndroidWorld and MobileWorld GUI-only evaluation runs for the base agents and their MobileForge-adapted variants. The repository is intended for result verification, log inspection, and mapping the public model checkpoints to the exact benchmark artifacts reported in the paper.

<!-- mobileforge-release-inventory:start -->
Release inventory `archives`: files=12; bytes=20015349616; sha256=b34dea1058bc00f8270e180e06b5d7707f6c9627f61e658dd738d2d8aad613de
<!-- mobileforge-release-inventory:end -->

## What is included?

- `archives/androidworld/*.tar.zst`: AndroidWorld rollout artifacts for Qwen3-VL-8B, GUI-Owl-1.5-8B, ForgeQwen3-8B, and ForgeOwl-8B.
- `archives/mobileworld/*.tar.zst`: MobileWorld GUI-only evaluation logs for the base and MobileForge-adapted models.
- `metadata/model_result_map.json`: machine-readable mapping from public archive names to public checkpoints, artifact sizes, hashes, and headline results.
- `metadata/archives_manifest.json`: compact manifest of all archive files and sizes.

The Hugging Face Dataset Viewer is disabled because the repository mainly stores compressed rollout logs rather than row-oriented examples.

## AndroidWorld results

AndroidWorld uses 116 tasks and reports Pass@1, Pass@2, and Pass@3.

| Setting | Checkpoint | Path | Pass@1 | Pass@2 | Pass@3 |
| --- | --- | --- | ---: | ---: | ---: |
| GUI-Owl-1.5-8B | `mPLUG/GUI-Owl-1.5-8B-Instruct` | `archives/androidworld/gui-owl-base.tar.zst` | 65/116 (56.0%) | 79/116 (68.1%) | 80/116 (69.0%) |
| ForgeOwl-8B, 200 tasks | `mobileforge-anonymous/ForgeOwl-8B-200tasks` | `archives/androidworld/gui-owl-mobileforge-200tasks.tar.zst` | 75/116 (64.7%) | 85/116 (73.3%) | 86/116 (74.1%) |
| ForgeOwl-8B, 400 tasks | `mobileforge-anonymous/ForgeOwl-8B-400tasks` | `archives/androidworld/gui-owl-mobileforge-400tasks.tar.zst` | 75/116 (64.7%) | 86/116 (74.1%) | 90/116 (77.6%) |
| ForgeOwl-8B, 900 tasks | `mobileforge-anonymous/ForgeOwl-8B` | `archives/androidworld/gui-owl-mobileforge-900tasks.tar.zst` | 78/116 (67.2%) | 87/116 (75.0%) | 90/116 (77.6%) |
| Qwen3-VL-8B | `Qwen/Qwen3-VL-8B-Instruct` | `archives/androidworld/qwen3-vl-base.tar.zst` | 47/116 (40.5%) | 57/116 (49.1%) | 64/116 (55.2%) |
| ForgeQwen3-8B, 200 tasks | `mobileforge-anonymous/ForgeQwen3-8B-200tasks` | `archives/androidworld/qwen3-vl-mobileforge-200tasks.tar.zst` | 55/116 (47.4%) | 64/116 (55.2%) | 71/116 (61.2%) |
| ForgeQwen3-8B, 400 tasks | `mobileforge-anonymous/ForgeQwen3-8B-400tasks` | `archives/androidworld/qwen3-vl-mobileforge-400tasks.tar.zst` | 61/116 (52.6%) | 69/116 (59.5%) | 73/116 (62.9%) |
| ForgeQwen3-8B, 900 tasks | `mobileforge-anonymous/ForgeQwen3-8B` | `archives/androidworld/qwen3-vl-mobileforge-900tasks.tar.zst` | 59/116 (50.9%) | 70/116 (60.3%) | 78/116 (67.2%) |

## MobileWorld GUI-only results

MobileWorld GUI-only uses 117 tasks. No MobileWorld rollout, task, or feedback is used for MobileForge adaptation.

| Setting | Checkpoint | Path | Success rate |
| --- | --- | --- | ---: |
| GUI-Owl-1.5-8B | `mPLUG/GUI-Owl-1.5-8B-Instruct` | `archives/mobileworld/gui-owl-base.tar.zst` | 44/117 (37.6%) |
| ForgeOwl-8B, 900 tasks | `mobileforge-anonymous/ForgeOwl-8B` | `archives/mobileworld/gui-owl-mobileforge-900tasks.tar.zst` | 48/117 (41.0%) |
| Qwen3-VL-8B | `Qwen/Qwen3-VL-8B-Instruct` | `archives/mobileworld/qwen3-vl-base.tar.zst` | 9/117 (7.7%) |
| ForgeQwen3-8B, 900 tasks | `mobileforge-anonymous/ForgeQwen3-8B` | `archives/mobileworld/qwen3-vl-mobileforge-900tasks.tar.zst` | 12/117 (10.3%) |

## Directory layout

```text
benchmark-results/
  metadata/
    archives_manifest.json
    model_result_map.json
  archives/
    androidworld/
      qwen3-vl-base.tar.zst
      qwen3-vl-mobileforge-200tasks.tar.zst
      qwen3-vl-mobileforge-400tasks.tar.zst
      qwen3-vl-mobileforge-900tasks.tar.zst
      gui-owl-base.tar.zst
      gui-owl-mobileforge-200tasks.tar.zst
      gui-owl-mobileforge-400tasks.tar.zst
      gui-owl-mobileforge-900tasks.tar.zst
    mobileworld/
      qwen3-vl-base.tar.zst
      qwen3-vl-mobileforge-900tasks.tar.zst
      gui-owl-base.tar.zst
      gui-owl-mobileforge-900tasks.tar.zst
```

Each archive expands to one experiment folder. AndroidWorld archives contain per-task compressed rollout artifacts (`*.pkl.gz`) and a run-level `*_results.json` summary. MobileWorld archives contain the corresponding MobileWorld evaluation logs and screenshots.

## Download examples

Download only the metadata:

```python
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="mobileforge-anonymous/mobileforge-benchmark-results",
    repo_type="dataset",
    filename="metadata/model_result_map.json",
)
```

Download one AndroidWorld run:

```python
from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="mobileforge-anonymous/mobileforge-benchmark-results",
    repo_type="dataset",
    allow_patterns=[
        "archives/androidworld/qwen3-vl-mobileforge-900tasks.tar.zst",
        "metadata/model_result_map.json",
        "metadata/archives_manifest.json",
    ],
)
```

Extract an archive:

```bash
tar --use-compress-program=unzstd -xf archives/androidworld/qwen3-vl-mobileforge-900tasks.tar.zst
```

## Notes

- AndroidWorld summaries can be parsed with the AndroidWorld checkpoint parser released in the MobileForge codebase.
- MobileWorld logs follow the upstream MobileWorld evaluation format.
- Some source rollout pickle members were already truncated and unusable. In the anonymous archives, each such member is represented by a valid status placeholder; run-level result summaries are unchanged.
- Compressed pickle files should be treated as executable Python serialization. Load them only in a trusted research environment.

## Citation

Citation information is withheld during double-blind review.