Fine-Tuning LLaMA-3.2 Instruct with QLoRA

Description

This repository contains a notebook for fine-tuning the meta-llama/Llama-3.2-3B-Instruct (or other generative language models) model using Quantized LoRA (QLoRA) for sentiment classification on the Arabic HARD dataset. This method allows efficient adaptation of large language models for specific tasks using low-rank approximations, reducing memory requirements and computational overhead.

Fine-Tuned Models

The fine-tuned models, available on Hugging Face, are as follows:

• YousraC/llama32_3B_ASA_HARD
• YousraC/llama321B_ASA_HARD

Note: These models were fine-tuned on a subset of the HARD dataset, not the entire dataset.

from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "YousraC/llama32_3B_ASA_HARD"
# Load the tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_name)
# Load the fine-tuned model
model = AutoModelForCausalLM.from_pretrained(model_name)

References

Fine-Tuning LLaMA with QLoRA

HARD Dataset

• Title: Hotel Arabic-Reviews Dataset Construction for Sentiment Analysis Applications
  Authors: A. Elnagar, Y. S. Khalifa, & A. Einea
  Publication: Procedia Computer Science, Volume 142, 2018, Pages 182-189
  DOI: 10.1016/j.procs.2018.10.466
  GitHub Repository: HARD Dataset GitHub

QLoRA

• Title: QLoRA: Efficient Finetuning of Quantized LLMs
  Authors: T. Dettmers, A. Pagnoni, A. Holtzman, & L. Zettlemoyer
  Publication: arXiv preprint, arXiv:2305.14314, 2023
  ArXiv Link: arxiv.org/abs/2305.14314
  GitHub Repository: QLoRA GitHub