From Fiction to Function: Leveraging Fine-Tuned Open-Domain Models for Character Mimicry

Overview

This repository contains a collection of experiments and analyses focused on the parameter-efficient fine-tuning of Large Language Models (LLMs) to mimic fictional characters. We use Spongebob as an example. The experiments are primarily conducted using Python and Jupyter Notebooks.

Repository Structure

Below is the structure of the repository with a brief description of each component:

• README.md: This file, providing an overview and instructions for the repository.
• data: Directory containing datasets used in the experiments.
• falcon_lora.ipynb: Jupyter Notebook for the Falcon LoRa experiment.
• gpt_finetune.ipynb: Jupyter Notebook for fine-tuning GPT models.
• graphing: Folder containing scripts and notebooks for graphing and visualizing data with custom formatting.
• outputs: Directory where output files from experiments are stored.
• peft.ipynb: Jupyter Notebook for experimenting with PEFT (Performance-Efficient Fine-Tuning).
• prefix_tuning.ipynb: Jupyter Notebook for prefix tuning experiments.
• prompt_tuning.ipynb: Jupyter Notebook for prompt tuning experiments.
• requirements.txt: File listing the necessary Python packages.
• scripts: Miscellaneous scripts used in various parts of the project.
• soft_prompt_tuning.ipynb: Jupyter Notebook for soft prompt tuning experiments.
• temp.py: Temporary Python script for auxiliary purposes.
• testing: Directory for testing scripts and experimental code.

Getting Started

Installation

To get started, clone the repository and install the required packages:

git clone [repository URL]
cd [repository name]
pip install -r requirements.txt

Running the Experiments

Each experiment is contained within its own Jupyter Notebook. To run an experiment:

1. Navigate to the notebook file (e.g., gpt_finetune.ipynb).
2. Open the notebook using Jupyter Lab or Jupyter Notebook.
3. Run the cells in the notebook sequentially.

These notebooks were run on a set of 8GB GPUs and should take a couple of hours to execute completely.

Validating the Results

After running the experiments, you can validate the results by comparing the generated graphs and outputs with those in the outputs directory. The graphs and output metrics should be comparable to ours with reasonable accuracy if the experiments are executed correctly. We generally saved our data and and re-graphed them using the scripts in the graphing library, but this was purely for formatting purposes. Given constraints, we were unable to save weights.

Support

For any queries or issues, please contact us at [email protected], or any of the other listed authors.

---