Fine Tuning with Databricks

This repository demonstrates how to fine-tune Large Language Models (LLMs) on Databricks using the HuggingFace Transformers Trainer framework. The project includes examples of standard fine-tuning as well as parameter-efficient fine-tuning with LoRA and QLoRA techniques for efficient model optimization.

---

Overview

The project guides users through the following phases:

1. Environment setup on Databricks.
2. Data import and preparation.
3. Model fine-tuning using the HuggingFace Trainer framework.
4. Parameter-efficient fine-tuning with LoRA.
5. Parameter-efficient fine-tuning with QLoRA.

---

Step-by-step Notebook Guide

1. 00.setup.ipynb — Initial Setup

• Create Unity Catalog Schema: Sets up the data space on Databricks to manage datasets and models.
• Create UC Volumes: Prepares storage areas for files and temporary data.

Note: File transfer to the Unity Catalog volume must be done manually through the Databricks interface or upload tools, following Databricks policies. Notebooks do not automate this operation.

2. 01.import_data.ipynb — Data Import and Preparation

• Load datasets: Imports JSONL files (train, validation, test) from UC volumes.
• Prepare data: Combines and transforms datasets, creates labels, and normalizes data.
• Save as Delta Table: Exports prepared data in Delta format, optimal for use in Databricks and ML workflows.

3. 02.fine_tuning_with_trainer.ipynb — Model Fine-tuning

• ML Environment Setup: Configures necessary libraries (Transformers, PyTorch, MLflow).
• Load data: Imports Delta datasets.
• Configure model: Sets the base model (e.g., BERT) and training parameters.
• Execute fine-tuning: Trains the model on prepared data using the HuggingFace Trainer framework.
• Evaluate and save: Measures performance and logs the model with MLflow.

4. 03.fine_tuning_with_lora.ipynb — LoRA Fine-tuning

• LoRA Configuration: Sets up parameter-efficient fine-tuning using LoRA (Low-Rank Adaptation).
• Reduced Parameters: Fine-tunes only a small subset of parameters while freezing the base model.
• Memory Efficient: Requires significantly less memory and computational resources than full fine-tuning.
• Load data: Imports Delta datasets.
• Execute training: Trains LoRA adapters on prepared data.
• Evaluate and save: Merges adapters with base model and logs to MLflow.

5. 04.fine_tuning_with_qlora.ipynb — QLoRA Fine-tuning

• QLoRA Configuration: Combines LoRA with 4-bit quantization for maximum efficiency.
• 4-bit Quantization: Uses bitsandbytes library to quantize the base model to 4-bit precision.
• Ultra Memory Efficient: Enables fine-tuning of large models on limited GPU resources.
• Load data: Imports Delta datasets.
• Execute training: Trains QLoRA adapters on quantized model.
• Evaluate and save: Merges adapters, dequantizes for inference, and logs to MLflow.

---

How to Use

1. Clone this repository on Databricks or locally.
2. Install dependencies listed in requirements.txt.
3. Follow the notebooks in order to learn and reproduce the fine-tuning workflow.

---

Requirements

• Databricks Account
• Python 3.8+
• HuggingFace Transformers, PyTorch, MLflow, and common ML libraries

Cluster Configuration

The fine-tuning notebooks are designed to run on a Databricks cluster with the following configuration:

• Spark Version: 15.4.x-cpu-ml-scala2.12
• Node Type: Standard_D16ds_v5 (16 cores, 64 GB memory)
• Driver Node Type: Standard_D16ds_v5
• Autotermination: 60 minutes
• Data Security Mode: SINGLE_USER
• Runtime Engine: STANDARD
• Cluster Mode: Single Node (0 workers)

For optimal performance, ensure the cluster has sufficient resources for model training.

---

Notes

These notebooks are intended for educational and experimental use. Adapt them to your needs before using in production.

---

For questions or suggestions, open an issue!