Fine-Tuning FLAN-T5 with LoRA

This repository contains a Jupyter Notebook for fine-tuning the FLAN-T5 Large model using LoRA (Low-Rank Adaptation) for sentiment classification.
The workflow demonstrates efficient fine-tuning with attention masks and padding to handle variable-length text data.

Main Libraries used:

• transformers
• torch
• scikit-learn
• pandas
• numpy

Description

The notebook covers:

• Importing libraries and setting up GPU/MPS acceleration if available.
• Data preparation: loading the dataset (Q2_20230202_majority 1.csv), preprocessing, label encoding, and splitting into train/test sets.
• Tokenization & Padding:
  • Tokenized input text with Hugging Face’s T5 tokenizer.
  • Applied attention masks to prevent tokens other than those associated with the labels to be used in the loss function
  • Used padding tokens for sequence alignment.
• LoRA Fine-Tuning:
  • Implemented custom LoRALayer and LinearWithLoRA modules.
  • Replaced linear layers in FLAN-T5 with LoRA-enabled layers (replace_linear_with_lora).
  • Froze all original model parameters — only LoRA parameters were trained.
• Training Loop:
  • Batched data using PyTorch DataLoader.
  • Forward pass with attention masks applied.
  • Optimized only LoRA parameters for efficient fine-tuning.
• Evaluation:
  • Model evaluated on the test split.
  • Accuracy tracked across epochs.

Results

The notebook demonstrates how LoRA drastically reduces the number of trainable parameters compared to full fine-tuning, while still adapting FLAN-T5 effectively.
It also shows how padding and attention masks help stabilize training with variable-length inputs and for classification respectively.