Ocg - Fine-Tuned Llama-3.1-8B Model

This repository contains a fine-tuned version of the meta-llama/Llama-3.1-8B model using LoRA (Low-Rank Adaptation) on an Apple MacBook Pro M3 with 36GB RAM, leveraging the MPS (Metal Performance Shaders) backend. The model was trained on a custom dataset (training_data.txt) to generate explanations or commands, and all necessary files are included for easy setup and deployment with Ollama.

Project Overview

• Base Model: meta-llama/Llama-3.1-8B
• Fine-Tuning Method: LoRA with r=16, lora_alpha=64, and target modules ["q_proj", "v_proj", "k_proj", "o_proj"]
• Hardware: MacBook Pro M3 with MPS acceleration
• Training Data: 91 examples in training_data.txt
• Date Completed: June 26, 2025

Files

• train_ocg.py: Python script used to fine-tune the model.
• fine_tuned_llama_m3/: Directory with fine-tuned LoRA adapter weights and tokenizer:
  • adapter_model.safetensors (54.5MB): Fine-tuned LoRA weights.
  • adapter_config.json: LoRA configuration.
  • tokenizer.json (17.2MB), tokenizer_config.json, special_tokens_map.json: Tokenizer files.
• training_data.txt: The dataset used for fine-tuning (91 examples).
• merge_model.py: Script to merge the adapter with the base model.

Setup Instructions

1. Prerequisites

• Operating System: macOS (tested on MacBook Pro M3).
• Python: Version 3.9 or higher.
• Homebrew: Install if not present (/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)").
• Ollama: Install via Homebrew (brew install ollama).

2. Create a Virtual Environment

• Set up a isolated Python environment:

  python3 -m venv ocg_env
  source ocg_env/bin/activate
  pip3 install transformers  peft  datasets  torch torchvision