zhongweixie/showui-worldmodel-results

ShowUI + WorldModel: Training Results & Artifacts

This dataset contains evaluation results, world models, and training data from the ShowUI + WorldModel integration project.

📦 Contents

1. Evaluation Results (results/miniwob_predictions/)

• Size: ~460MB
• Format: JSONL files with episode-level predictions
• Tasks: 9 MiniWoB++ tasks evaluated with ShowUI agent
• Includes:
  • Task success/failure outcomes
  • Action predictions and execution traces
  • World model predictions (when enabled)
  • Screenshots and DOM snapshots

2. MiniWoB Metadata (datasets/miniwob_metadata/)

• Size: ~134MB
• Format: JSON files with task metadata
• Files:
  • hf_train.json - Training set metadata (9,829 tasks)
  • hf_test.json - Test set metadata
• Includes:
  • Task instructions and goals
  • Initial state descriptions
  • Screenshots (base64 encoded)
  • DOM structure

3. PatchWorld Trajectories (data/patchworld_trajectories/)

• Size: ~996KB
• Format: JSONL files in PatchWorld format
• Tasks: 9 MiniWoB++ tasks
• Includes:
  • State transitions with before/after observations
  • Action specifications
  • Success indicators
  • Split into train/test sets

4. Induced World Models (data/induced_world_models/)

• Size: ~20KB
• Format: Python modules
• Models:
  • choose-date_world_model.py - Calendar date selection
  • grid-coordinate_world_model.py - Grid navigation
  • count-shape_world_model.py - Shape counting
• Induced by: Claude Opus 3.7 via PatchWorld inducer

🎯 Use Cases

1. Reproduce Evaluations: Use the prediction results to analyze ShowUI performance
2. World Model Research: Study the induced world models for abstract task understanding
3. Training Data: Use PatchWorld trajectories for imitation learning
4. Metadata Analysis: Explore MiniWoB task structure and difficulty

📊 Key Findings

World Model Impact (Preliminary)

• choose-date: World model correctly predicts calendar behavior
• grid-coordinate: Spatial reasoning for grid navigation
• count-shape: Accurate shape counting before action

Training Configurations

• gc_fix: Gradient accumulation fix for stability
• stratified: Balanced sampling across task difficulties

🔗 Related Resources

• Code Repository: https://github.com/zhongweixie/gui_patch_world
• ShowUI Paper: arXiv:2411.17465
• PatchWorld Framework: State transition modeling for GUI agents

📝 Citation

If you use this dataset, please cite:

@misc{showui-worldmodel-2024,
  title={ShowUI + WorldModel: Integrating World Models for GUI Agent Training},
  author={Xie, Zhongwei},
  year={2024},
  howpublished={HuggingFace Dataset: zhongweixie/showui-worldmodel-results}
}

@article{lin2024showui,
  title={ShowUI: One Vision-Language-Action Model for GUI Visual Agent},
  author={Lin, Kevin Qinghong and others},
  journal={arXiv preprint arXiv:2411.17465},
  year={2024}
}

🛠️ Technical Details

Evaluation Framework

• Environment: MiniWoB++ (Chrome-based)
• Agent: ShowUI-2B with LoRA adapters
• World Model: PatchWorld-induced Python functions
• Metrics: Task success rate, action accuracy

Data Processing

• Trajectory conversion: convert_to_patchworld.py
• World model induction: run_inducer.sh
• Evaluation: eval_online.py with worldmodel_agent.py

Training Setup

• Base model: ShowUI-2B
• Training method: LoRA fine-tuning
• Batch size: 1 per GPU (gradient accumulation: 8)
• Learning rate: 1e-4
• Epochs: 3

📧 Contact

For questions or issues, please open an issue on the GitHub repository.

---

License: Same as ShowUI (Apache 2.0)
Last Updated: 2024-08-12