Model Fine-Tuning on CPU

This project demonstrates how to fine-tune a pre-trained transformer model on a CPU using the Hugging Face transformers, peft, and datasets libraries, along with PyTorch. It leverages Parameter-Efficient Fine-Tuning (PEFT) with LoRA (Low-Rank Adaptation) to optimize the fine-tuning process for resource-constrained environments.

Prerequisites

• Python 3.8 or higher
• A CPU-based system (no GPU required)
• Git installed for version control

Installation

1. Clone this repository:

   git clone https://github.com/choosechart/fine-tune-cpu.git
   cd fine-tune-cpu

2. Install dependencies using pip:

   pip install transformers peft datasets torch

Alternatively, use uv to manage dependencies:

uv sync

Project Structure

• fine_tune.py: Main script containing the fine-tuning logic.
• README.md: This file, providing project overview and instructions.
• .gitignore: Excludes unnecessary files (e.g., virtual environments, model checkpoints).

Dependencies

The project relies on the following Python libraries:

• transformers: Provides pre-trained models and tokenizers for natural language processing tasks.
  • AutoModelForSequenceClassification: Loads a pre-trained model for sequence classification tasks.
  • AutoTokenizer: Handles tokenization of input text.
  • Trainer: Manages the training loop for fine-tuning.
  • TrainingArguments: Configures training hyperparameters.
• peft: Enables parameter-efficient fine-tuning.
  • LoraConfig: Configures LoRA for efficient model adaptation.
  • get_peft_model: Applies LoRA to the pre-trained model.
• datasets: Manages datasets for training and evaluation.
  • Dataset: Represents a single dataset.
  • DatasetDict: Organizes multiple datasets (e.g., train, validation, test).
• torch: PyTorch library for tensor operations and model training on CPU.

Usage

1. Prepare your dataset in a format compatible with the datasets library (e.g., a DatasetDict with train and validation splits).
2. Configure the model and LoRA settings in main.py.
3. Run the fine-tuning script:

   uv run python fine_tune.py --output_dir "./fine_tuned_distilbert" --model_name "distilbert/distilbert-base-uncased" --data_file_path "data_dict_example.json"

   Or, if using uv:

   uv run python fine_tune.py --output_dir "./fine_tuned_distilbert" --model_name "distilbert/distilbert-base-uncased" --data_file_path "data_dict_example.json"

Notes

• This project is optimized for CPU usage, making it suitable for environments without GPU support.
• LoRA reduces memory and computational requirements, enabling efficient fine-tuning on modest hardware.
• Ensure sufficient memory (at least 16GB RAM recommended) for handling large transformer models.

License

This project is licensed under the MIT License.