CP-Fuse: Model Fusion for Copyright-Protected Language Generation

• Overview
• Contents
• Getting Started
• Usage
• Contributing
• Contact
• Citation

Overview

This repository contains the Python implementation of Copyright-Protecting Model Fusion (CP-Fuse), a novel method for copyright protection introduced in the ICLR 2025 Oral paper "Copyright-Protected Language Generation via Adaptive Model Fusion".

Traditional methods for mitigating copyright risks often involve post-processing constraints applied at inference time. However, these methods can be computationally intensive and may compromise the model's performance. CP-Fuse introduces a novel strategy that fuses the outputs of multiple models in a way that protects against copyright infringement. In particular, CP-Fuse adaptively aggregates the logits to minimize the reproduction of copyrighted content, adhering to a crucial balancing property to prevent the regurgitation of memorized data.

Contents

The cp_fuse folder contains the core code of the package, organized into several subfolders.

• cp_fuse/: Core package containing the implementation of CP-Fuse.
• examples/: Example scripts to demonstrate the usage of training, evaluation, and data processing functions.

Getting Started

Dependencies

• Python 3.12.3
• torch 2.3.0
• peft 0.10.0
• transformers 4.40.0
• datasets 2.19.1
• pynvml 11.5.0
• numpy 1.26.4
• pandas 2.2.2
• accelerate 0.30.1
• huggingface-hub 0.23.4
• bitsandbytes 0.44.1

Installation

To set up your environment and install the package, follow these steps:

Create and Activate a Conda Environment

Start by creating a Conda environment with Python 3.12.3. This step ensures your package runs in an environment with the correct Python version.

conda create -n cp_fuse python=3.12.3
conda activate cp_fuse

Install the Package

There are two ways to install the package:

1. Local Installation: Start by cloning the repository from GitHub. Then, upgrade pip to its latest version and use the local setup files to install the package.

   git clone https://github.com/jaabmar/cp_fuse.git
   cd cp_fuse
   pip install --upgrade pip
   pip install -e .

2. Direct Installation from GitHub: You can also install the package directly from GitHub.

   pip install git+https://github.com/jaabmar/cp_fuse.git

Usage

Training

The training script is accessible via the entry point train, which can be run directly from the terminal after the package is installed. This entry point corresponds to the examples/train.py script and includes flexible command-line arguments to specify model configuration, dataset splits, and training parameters.

Example Command:

train \
    --model_checkpoint meta-llama/Llama-2-7b-hf \
    --dataset_name StoryTelling \
    --epochs 20 \
    --seq_length 2048 \
    --n_train_samples 9000 \
    --num_splits 2 \
    --split 1 \
    --batch_size 1 \
    --learning_rate 5e-6 \
    --output_dir ./models

Training Arguments

• --model_checkpoint (str): Path to the model checkpoint. Default is meta-llama/Llama-2-7b-hf.
• --dataset_name (str): Name of the dataset to use for training.
• --epochs (int): Number of training epochs.
• --seq_length (int): Sequence length for model input.
• --n_train_samples (int): Number of training samples.
• --num_splits (int): Number of dataset splits for training. Default is 2.
• --split (int): Specify the training split (e.g., 1).
• --batch_size (int): Batch size per training step.
• --learning_rate (float): Learning rate for training.
• --quantize (flag): Enable 8-bit quantization (LoRA) if provided.
• --seed (int): Random seed for reproducibility.
• --gradient_accumulation_steps (int): Number of steps to accumulate gradients.
• --save_freq (int): Frequency (in epochs) to save model checkpoints.
• --eval_freq (int): Frequency (in epochs) to evaluate the model.
• --output_dir (str): Directory to save model checkpoints.

Evaluation

The evaluation script is accessible via the entry point evaluate. This entry point corresponds to the examples/evaluate.py script and evaluates trained models on specified datasets, using either a single model or combining two models with CP-Fuse.

Example Command:

evaluate \
    --model_checkpoint1 ./models/StoryTelling_split_1 \
    --model_checkpoint2 ./models/StoryTelling_split_2 \
    --dataset_name StoryTelling \
    --n_test_samples 500 \
    --batch_size 16 \
    --grid_size 10 \
    --output_dir ./eval

Evaluation Arguments

• --model_checkpoint1 (str): Path to the first model checkpoint. Required.
• --model_checkpoint2 (str): Path to the second model checkpoint (for CPModel).
• --dataset_name (str): Name of the dataset for evaluation.
• --n_test_samples (int): Number of test samples. Default is 500.
• --batch_size (int): Batch size for evaluation. Default is 32.
• --grid_size (int): Grid size for grid search when using CPModel. Default is 10.
• --verbose (flag): Enable verbose output during evaluation if provided.
• --output_dir (str): Directory to save evaluation results.

Contributing

We welcome contributions to improve this project. Here's how you can contribute:

1. Fork the Project
2. Create your Feature Branch (git checkout -b feature/AmazingFeature)
3. Commit your Changes (git commit -m 'Add some AmazingFeature')
4. Push to the Branch (git push origin feature/AmazingFeature)
5. Open a Pull Request

Contact

For any inquiries, please reach out:

• Javier Abad Martinez - [email protected]
• Konstantin Donhauser - [email protected]

Citation

If you find this code useful, please consider citing our paper:

@article{abad2025copyright,
     title={Copyright-Protected Language Generation via Adaptive Model Fusion}, 
     author={Javier Abad and Konstantin Donhauser and Francesco Pinto and Fanny Yang},
     journal={International Conference on Learning Representations},
     year={2025},
}