---
title: FFG Mask Explorer
emoji: 🔬
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 5.0.0
app_file: app.py
pinned: false
license: apache-2.0
hardware: a10g-small
---

# FFG Mask Explorer 🔬

An interactive tool for generating and visualizing Fast Fisher Grafting (FFG) masks on fine-tuned language models.

## Features

- **Real-time mask generation** using GPU acceleration
- **Multiple grafting methods**: FFG, Magnitude, and Fish-Mask
- **Interactive visualizations** showing sparsity patterns and statistics
- **Pre-configured models** from the paper's experiments
- **Custom model support** for your own fine-tuned models

## How to Use

1. **Select a Model**: Choose from pre-configured models or enter custom model IDs
2. **Set Sparsity**: Adjust the sparsity ratio (fraction of parameters to keep)
3. **Choose Method**: Select between FFG, Magnitude, or Fish-Mask grafting
4. **Generate**: Click to create masks and visualizations in real-time

## About FFG

Fast Fisher Grafting (FFG) uses the second moments from Adam optimizer to identify important parameters in fine-tuned models. This provides more informed pruning compared to magnitude-based methods.

Based on the paper: [Harnessing Optimization Dynamics for Curvature-Informed Model Merging](https://arxiv.org/abs/2509.11167)

## Technical Details

- **GPU**: Requires GPU for efficient processing (A10G recommended)
- **Memory**: ~24GB GPU memory for 8B parameter models
- **Models**: Compatible with Llama-3.1-8B based fine-tunes

## Local Development

To run this Space locally:

```bash
# Clone the repository
git clone https://huggingface.co/spaces/YOUR_USERNAME/ffg-mask-explorer

# Install dependencies
pip install -r requirements.txt

# Copy FFG experiment suite (adjust path as needed)
cp -r /path/to/surgeon/ffg_experiment_suite .

# Run the app
python app.py
```

## Citation

```bibtex
@misc{mahdavinia2025harnessingoptimizationdynamicscurvatureinformed,
    title={Harnessing Optimization Dynamics for Curvature-Informed Model Merging}, 
    author={Pouria Mahdavinia and Hamed Mahdavi and Niloofar Mireshghallah and Mehrdad Mahdavi},
    year={2025},
    eprint={2509.11167},
    archivePrefix={arXiv},
    primaryClass={cs.LG},
    url={https://arxiv.org/abs/2509.11167}, 
}
```

## License

Apache 2.0 # Force rebuild Sun Sep 28 19:47:52 EDT 2025