Gemma-Model-Finetuning-Using-Lora

Gemma-Model-Finetuning-Using-Lora

Fine tuning Domain Specific dataset (Personal Dataset) on Gemma 2B Model



This Model is Finetuned on a single P100 GPU , which is Freely available in Kaggle

This repository contains the code and data for fine-tuning a small language model (SLM) for the specific domain of Indian history. The project demonstrates how to adapt a pre-trained language model to better understand and generate text relevant to this historical context.

Requirements

• Kaggle/collab Account
• GPU access in Kaggle/collab
• Hugging Face API Key

Project Overview

Large language models (LLMs) have shown remarkable capabilities in natural language processing tasks. However, fine-tuning them for specific domains remains crucial to unlock their full potential. This project explores the adaptation of the GEMMA model for analyzing Indian history, utilizing a dedicated Indian history dataset. We employ techniques like BitsAndBytes quantization and LoraConfig customization to optimize the model for causal language modeling tasks within this domain.

Usage:

• clone the repository

git clone https://github.com/AjayK47/Gemma-Model-Finetuning-Using-Lora.git

• Run the Jupyter Notebooks in the following order:
• • dataset-preprocessing.ipynb
  • gemma-finetuned-model-inference.ipynb
  • gemma-it-finetuned.ipynb
• You can Try Out My Hugging Face Model from here - https://huggingface.co/Ajayk/indian-history-gemma-instruction-finetuned

Dependencies

Install the dependencies using the following command:

Methodology

The project follows these key steps:

• Data Preprocessing: Select and clean a specific Domain dataset. Format the data into the GEMMA model chat template.
• Fine-tuning:
  • Use the SFTTrainer from the Hugging Face Transformers library.
  • Incorporate domain specific parameters for causal language modeling.
  • Optimize training parameters (e.g., learning rate, epochs).
  • Integrate Per-sample Feature Transform (Peft) for enhanced adaptability.
• Evaluation and Deployment:
  • Evaluate the fine-tuned model on relevant tasks (e.g., text summarization, question answering).
  • Save and Push the model to Hugging Face.

Future Work

• Potential future directions include:
• • Expanding the approach to other domains and languages.
  • Exploring alternative compact language models.
  • Developing techniques to understand the model's reasoning and decision-making processes.

Contribution

• Contributions to this project are welcome!

License

This project is licensed under the MIT License.