README.md

---
license: apache-2.0
language:
- en
pretty_name: MobileForge Training Data
tags:
- mobileforge
- mobile-gui-agent
- grpo
- reinforcement-learning
- step-level-reward
- androidworld
- verl
task_categories:
- reinforcement-learning
- text-generation
viewer: false
---

# MobileForge Training Data

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

This dataset contains the hint-contextualized step-level GRPO training data used by MobileForge. The samples are produced from MobileGym rollouts, hierarchical critic feedback, and HiFPO filtering.

The Hugging Face Dataset Viewer is disabled because the release contains very large JSON-array files with multimodal conversation samples and rollout metadata.

<!-- mobileforge-release-inventory:start -->
Release inventory `payloads`: files=7; bytes=114495299717; sha256=00076b77cf46fc2f049ec19a5494a3bae58936a83390c76d1a09a0ee5ab3a9fb
<!-- mobileforge-release-inventory:end -->

## Files

| Split | File | Size | Description |
| --- | --- | ---: | --- |
| `qwen3-vl-900tasks` | `mobileforge_grpo_20260313_081857_tasks_900.json` | 39.25 GB | Qwen3-VL-8B MobileForge 900-task training split. |
| `qwen3-vl-400tasks` | `mobileforge_grpo_20260313_081857_tasks_400.json` | 17.13 GB | Qwen3-VL-8B MobileForge 400-task scaling split. |
| `qwen3-vl-200tasks` | `mobileforge_grpo_20260313_081857_tasks_200.json` | 8.32 GB | Qwen3-VL-8B MobileForge 200-task scaling split. |
| `gui-owl-900tasks` | `mobileforge_grpo_20260508_102018_tasks_900.json` | 24.60 GB | GUI-Owl-1.5-8B MobileForge 900-task training split. |
| `gui-owl-400tasks` | `mobileforge_grpo_20260508_102018_tasks_400.json` | 11.41 GB | GUI-Owl-1.5-8B MobileForge 400-task scaling split. |
| `gui-owl-200tasks` | `mobileforge_grpo_20260329_093821_tasks_200.json` | 5.80 GB | GUI-Owl-1.5-8B MobileForge 200-task scaling split. |
| `validation` | `mobileforge_grpo_20260307_021142_h3.json` | 128.51 MB | Validation set used during MobileForge GRPO training. |

`manifest.json` provides the same file list in machine-readable form.

## Data format

Each file is a JSON array. Each item is one step-level training example in the MobileForge conversation format.

Top-level fields:

| Field | Description |
| --- | --- |
| `conversations` | Multimodal chat messages consumed by the Qwen3-VL / GUI-Owl training stack. |
| `bad_step` | Whether the sample is marked as a bad step by filtering. |
| `rollout_log_probs_sum` | Sum of rollout log probabilities for the selected action response. |
| `rollout_logprobs` | Token-level rollout log probabilities. |
| `metadata` | Task id, attempt id, step number, feedback labels, final outcome, selected action, and other rollout/evaluator metadata. |

Common `metadata` fields include `task_id`, `attempt_id`, `step_number`, `step_label`, `reasonableness`, `impact`, `step_success`, `overall_success`, `final_result`, `action_type`, and `action`.

## Download examples

Download only the manifest and validation split:

```python
from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="mobileforge-anonymous/mobileforge-training-data",
    repo_type="dataset",
    allow_patterns=[
        "manifest.json",
        "mobileforge_grpo_20260307_021142_h3.json",
    ],
)
```

Download the 900-task Qwen3-VL split:

```python
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="mobileforge-anonymous/mobileforge-training-data",
    repo_type="dataset",
    filename="mobileforge_grpo_20260313_081857_tasks_900.json",
)
```

## Inspect large JSON files safely

These files are large. Avoid `json.load` for the training splits unless you have enough memory. Use a streaming parser such as `ijson`:

```python
import ijson

with open("mobileforge_grpo_20260313_081857_tasks_900.json", "rb") as f:
    for example in ijson.items(f, "item"):
        print(example.keys())
        print(example["metadata"])
        break
```

## Training usage

The public MobileForge code release includes a training entry point compatible with this data:

```bash
bash training/examples/qwen3_vl_8b_mobileforge_grpo.sh \
  --model_path Qwen/Qwen3-VL-8B-Instruct \
  --data_path /path/to/mobileforge_grpo_20260313_081857_tasks_900.json \
  --val_data_path /path/to/mobileforge_grpo_20260307_021142_h3.json \
  --experiment_name qwen3_vl_8b_mobileforge_grpo_900tasks \
  --total_epochs 4 \
  --reward_action_type_weight 0.2 \
  --reward_action_params_weight 0.8
```

## Relationship to other MobileForge artifacts

This dataset is generated from:

1. `mobileforge-anonymous/mobileforge-generated-tasks`: automatically generated AndroidWorld-side tasks.
2. `mobileforge-anonymous/mobileforge-exploration-trajectories`: source exploration traces used for task generation.
3. MobileGym rollout and critic feedback produced by the MobileForge codebase.

The resulting checkpoints are released on the [anonymous Hugging Face profile](https://huggingface.co/mobileforge-anonymous), and their benchmark logs are released in `mobileforge-anonymous/mobileforge-benchmark-results`.

## Limitations

- The examples are specialized for MobileForge-style mobile GUI action prediction and step-level GRPO.
- The files may contain screenshots and environment traces from Android app interaction. Use them for research and reproducibility.
- The JSON arrays are intentionally kept close to the training format; they are not optimized for interactive browsing.

## Citation

Citation information is withheld during double-blind review.