README.md

Parameter Efficient Fine-Tuning

Overview

This directory provides demonstrations of various Parameter Efficient Fine-Tuning (PEFT) techniques.

Contents

• PEFT Demo (peft_demo.ipynb): A demo on using PEFT techniques like LoRA and DORA, and their implementations, with functions to be completed by participants.
• Custom Lightning Module (custom_lightning_module.py): A Python module that defines custom Lightning components used in the notebook. This module is crucial for the implementation of the custom LoRA layer described in the tutorial.

Current Demos

PEFT Tutorial

• Filename: peft_demo.ipynb
• Objective: To fine-tune a model using PEFT methods, with an emphasis on practical implementations using custom LoRA layers and exploration of DORA technique.
• Key Features:
  • Detailed explanation and implementation of a custom LoRA layer.
  • Uses the custom LoRA layer to fine-tune a DistilBERT model on the IMDb dataset, from the ../../data/imdb directory.
  • Demonstrates the application of PEFT techniques using the HuggingFace peft library.
  • Provides a showcase on DORA.
  • Contains incomplete functions indicated by "TODO" comments which participants are expected to implement as part of the learning process.