I Learn Better If You Speak My Language: Understanding the Superior Performance of Fine-Tuning Large Language Models with LLM-Generated Responses

Overview

This project includes training scripts, evaluation tools, and datasets for the paper I Learn Better If You Speak My Language: Understanding the Superior Performance of Fine-Tuning Large Language Models with LLM-Generated Responses. It builds upon the LLAMA-FACTORY project to train and test language models effectively.

Features

• Training Scripts: Customizable training scripts leveraging LLAMA-FACTORY.
• Evaluation Tools: Unified prediction functions for seamless evaluation.
• Datasets Included: All necessary datasets are provided within the project.
• Perplexity Calculation: (Coming soon) Scripts for perplexity calculations will be added.

Getting Started

1. Prerequisites

   • Python 3.10

2. Set Up LLAMA-FACTORY Directory

   Modify the LLAMA_FACTORY_DIRECTORY_new variable in your scripts to point to your LLAMA-FACTORY directory:

   LLAMA_FACTORY_DIRECTORY_new = '/path/to/your/LLAMA-FACTORY'

3. Install Dependencies

   Install the required Python packages:

   pip install -r requirements.txt

Training

To train the model:

1. Modify the Training Script

   In utils/train.py, locate the train_llama_factory function and update the model path:

Evaluation

To evaluate the trained model:

1. Modify the Evaluation Script

   In evaluation/eval.py, find the do_predict_llama_factory_unify function and set your model path:

Datasets

All datasets required for training and evaluation are available in the datasets/ directory.

Perplexity Calculation

Note: The perplexity calculation scripts will be added soon.

Recommended Workflow

For the best experience:

1. Create Custom Training Scripts

   • Start by creating your own training scripts based on the provided templates.
   • Customize them according to your project's needs.

2. Train Your Model

   • Use the modified training scripts to train your model.
   • Ensure all paths and configurations point to your directories and models.

3. Evaluate the Model

   • After training, use the evaluation scripts to assess your model's performance.
   • Merge or adapt the evaluation scripts into your project as needed.

Acknowledgments

• LLAMA-FACTORY: This project builds upon the excellent work done in the LLAMA-FACTORY project.