Llama-3.2-3B-Instruct-bnb-4bit

Greek Translation Model - Fine-tuning and Evaluation

This repository contains two Jupyter notebooks designed for fine-tuning and evaluating a Greek language model using state-of-the-art tools and optimizations.

Contents

1. NLP_FINETUNE.ipynb - A notebook for fine-tuning the Greek language model.
2. NLP_Eval.ipynb - A notebook for evaluating the fine-tuned model on test data.

Requirements

• Python: 3.8+
• Libraries: The notebooks automatically handle the installation of required libraries, but ensure that your environment includes:
  • unsloth
  • xformers
  • trl
  • peft
  • accelerate
  • bitsandbytes
  • triton
  • datasets
  • transformers
  • torch (PyTorch)

Fine-tuning Process (NLP_FINETUNE.ipynb)

1. Install Dependencies: The notebook installs all required libraries to ensure efficient fine-tuning.
2. Load Dataset: Leverages Hugging Face's datasets to load the dataset for fine-tuning.
3. Initialize Trainer: Uses SFTTrainer from the trl library for supervised fine-tuning.
4. Model Optimization: The model is optimized using FastLanguageModel from unsloth, with support for bfloat16 (if available).

Evaluation Process (NLP_Eval.ipynb)

1. Install Dependencies: Similar to the fine-tuning notebook, all necessary packages are installed.
2. Load Model and Dataset: Loads the fine-tuned model and prepares the test dataset for evaluation.
3. Evaluation Metrics: The model is evaluated using standard language model metrics.

How to Use

1. Run the Fine-tuning Notebook:

   • Open NLP_FINETUNE.ipynb in a Jupyter environment and run all cells to fine-tune the model.

2. Evaluate the Model:

   • After fine-tuning, use the NLP_Eval.ipynb notebook to evaluate the model's performance.

Acknowledgements

The project uses cutting-edge tools from the Hugging Face ecosystem and optimizations from the unsloth library for rapid and efficient fine-tuning.