This repository includes the steps I took to fine-tune a large language model (LLM) for sentiment analysis using the IMDB dataset. The process leverages Hugging Face's transformers and datasets libraries, as well as the peft library for parameter-efficient fine-tuning using LoRA (Low-Rank Adaptation).
The fine-tuned sentiment analysis model can be found on Hugging Face at: samadpls/sentiment-analysis.
We use the distilbert-base-uncased model as the base model for fine-tuning. The model is configured for sequence classification with two labels (positive and negative sentiment).
LoRA (Low-Rank Adaptation) is used to fine-tune the model efficiently by adding trainable low-rank adaptation matrices to certain model layers. The configuration parameters for LoRA used in this project are:
By leveraging LoRA, we achieve efficient fine-tuning with reduced computational resources, making it a key component of this project.
The model's performance is evaluated using accuracy as the metric. The trained model is tested on a set of example texts to verify its predictions.
I developed this project in 2023 while learning to run and fine-tune LLMs. This project serves as a starting point for fine-tuning LLMs. Using LoRA, we can efficiently fine-tune large models with reduced computational resources. Feel free to experiment with various models, datasets, and configurations to enhance your understanding and achieve better results.