Parameter Efficient Fine-Tuning (PEFT) Plugin Framework

Project Overview

In Artificial Intelligence (AI), Foundation Models such as Large Language Models (LLMs) are large-scale machine learning models trained on Text datasets for NLP and NLG requirements. The models are then adapted with fine-tuning for a wide variety of downstream NLP applications such as Classification, Content Generation, Language translation, Information searching and conversational AI. However, as LLMs grow in scale, fine-tuning them on downstream tasks becomes computationally and memory-intensive, as fine-tuning is performed entirely on a pre-trained model with new data.

Parameter-Efficient Fine-Tuning (PEFT) techniques are a set of methods that perform fine-tuning on only a small subset of the parameters of a pre-trained model, such as an LLM, while achieving desired performance with reduced computational requirements.

Project Objectives and Scope

1. Perform a high-level theoretical study on PEFT techniques, including their functionalities, advantages, and limitations.
2. Design and implement API-driven PEFT functional modules that can be consumed during fine-tuning ML Models.
3. Leverage publicly available PEFT techniques, covering LoRA and QLoRA.
4. Design and implement a basic plugin framework that enables integrating Modularized PEFT techniques for consumption by LLMs.
5. Implement a basic integration of any one open-source Language Model (e.g., BERT) to the Plugin Framework to consume the PEFT techniques.
6. Demonstrate the Plugin framework capabilities with a minimal user experience.
7. Implement a basic graphical UI using Streamlit.

Project Team

• Sanjana C
• Tharun
• Pooja Kulkarni
• Sujal Singh

Project Deliverables

1. A theoretical study on PEFT techniques, including their functionalities, advantages, and limitations.
2. A plugin framework for integrating PEFT techniques, with a focus on LoRA and QLoRA.
3. A basic integration of an open-source Language Model (e.g., BERT) to the Plugin Framework.
4. A minimal user experience demonstrating the Plugin framework capabilities.
5. A basic graphical UI using Streamlit.

OOD Design of the Project

Comparison of PEFT Techniques

Project Presentation

The project presentation can be found (PRESENTATION_LINK).

Video Tutorial

Vision_Model Video

Vision_Model.-.Made.with.Clipchamp.mp4

Final output over the UI

Bert_Model Video

BERT_Model.-.Made.with.Clipchamp.mp4

Final output over the UI