Full Fine-Tuning of Flan-t5-base Model

This project demonstrates the process of fully fine-tuning the Flan-T5-Base model for the NVIDIA question-answering task. The main objective of this project is to provide beginners with hands-on experience in fine-tuning a large language model, rather than achieving a perfect model.

Environment Setup

1. Setup Conda environment

conda create -n fine-tuning-q-and-a python=3.11
conda activate fine-tuning-q-and-a

2. Install dependencies

pip install -r requirements.txt

Overview

• The Flan-T5-Base model, a powerful language model, is used as the base model for fine-tuning.
• The model is fine-tuned specifically for the NVIDIA question-answering task.
• The project serves as an educational resource for beginners to understand and practice the fine-tuning process.

Machine Learning Process

• The dataset used for fine-tuning the model consists of question-answer pairs related to NVIDIA taken from Kaggle. The dataset is prepared appropriately for training the model followed by performing tokenization on the whole dataset. Check 01_Data_Preparation.ipynb and 02_Data_Tokenization.ipynbfor running the code implementation.
• Model training was done using the tokenized dataset in 03_Model_Training.ipynb notebook.
• Once training was done, model was then evaluated based on performance metric - evaluation loss and also based on qualitative analysis in the notebooks - 04_Model_Evaluation_Performance_Metric.ipynb and 05_Model_Evaluation_Qualitative_Analysis.ipynb.
• For nice readibility of the code, all of the functions are included in helper.py and the environment variables are loaded constants.py in utils folder. Check these files for more details understanding of the functions.

Reference

Special thanks to Eng. Omar M. Atef for creating course on Udemy: LLMs Workshop: Practical Exercises of Large Language Models. Full video tutorial can be found in Section 2 of this course.