How To Fine-tune LLM for Arabic Instructions Using LoRA

In this article, we’ll dive deep into the process of fine-tuning a large language model (LLM) using Low-Rank Adaptation (LoRA). We’ll use a Qwen1.5–7B model on an Arabic instruction dataset.

Image By Author

Let’s break down each section of the code and explore its purpose and functionality!

1. Importing Libraries:

from datasets import load_dataset
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, TrainingArguments
from peft import LoraConfig, get_peft_model
from trl import SFTTrainer

This section imports the required libraries:
- datasets: For loading and managing datasets
- torch: The PyTorch deep learning framework
- transformers: Hugging Face’s library for working with pre-trained models
- peft: Parameter-Efficient Fine-Tuning library
- trl: TRL library for reinforcement learning and supervised fine-tuning

2. Loading the Dataset

This dataset contains six million instruction-response pairs in Arabic, which will be used to fine-tune our model.

dataset_name = 'akbargherbal/six_millions_instruction_dataset_for_arabic_llm_ft'
dataset = load_dataset(dataset_name, split="train")

3. Defining and Loading the Base Model

model_name = "Alibaba-NLP/gte-Qwen1.5–7B-instruct"
model = AutoModelForCausalLM.from_pretrained(
 model_name,
 quantization_config=bnb_config,
 trust_remote_code=True
)
model.config.use_cache = False

4. Configuring Quantization Settings

bnb_config = BitsAndBytesConfig(
 load_in_4bit=True,
 bnb_4bit_quant_type="nf4",
 bnb_4bit_compute_dtype=torch.float16,
)

To reduce memory usage and increase training speed, we use 4-bit quantization. This configuration uses the “normal float 4” (nf4) quantization type and sets the compute dtype to float16 for a good balance between precision and efficiency.

5. Setting up LoRA (Low-Rank Adaptation)