llama3-8b-dpo-orca-adapter

🦙 Llama-3 DPO Alignment Pipeline

A professional, modular pipeline to fine-tune Llama-3 8B using Direct Preference Optimization (DPO) — from training to inference to benchmarking.

---

🚀 What is This Project?

This repository contains a complete, production-grade pipeline for aligning a large language model using DPO. Instead of manually patching notebooks together, this project is structured as reusable Python scripts with YAML configurations so anyone can reproduce the training with a single command.

⚡ The Core Idea

---

🗂️ Project Structure

llama3-dpo-alignment-pipeline/
├── 📁 configs/
│   ├── dpo_config.yaml          # All DPO training hyperparameters
│   └── benchmark_config.yaml   # Test prompts & generation settings
├── 📁 scripts/
│   └── train_dpo.py            # The main training engine (reads from configs/)
├── 📁 inference/
│   └── inference.py            # Load adapter and run interactive inference
├── 📁 evaluation/
│   └── benchmark.py            # Compare Base vs. Aligned model side-by-side
├── 📁 training/
│   └── training-llama3-dpo.ipynb  # The original Kaggle notebook
├── 📁 models/                  # (gitignored) Local adapter weights live here
├── pyproject.toml              # Dependency management with uv
└── README.md

---

🛠️ Setup & Installation

This project uses uv — the fastest Python package manager. Everything is managed in one command.

1. Install uv

# Windows
powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

2. Clone the Repo & Sync Dependencies

git clone https://github.com/Edge-Explorer/llama3-dpo-alignment-pipeline.git
cd llama3-dpo-alignment-pipeline
uv sync

Note: Unsloth requires an NVIDIA GPU to import. For local development, you can write and review code without a GPU. Use Kaggle or Google Colab to actually run the scripts.

---

🏋️ Training

All training parameters are in configs/dpo_config.yaml. You can tweak learning rates, batch sizes, and sequence lengths without touching any Python code.

# On Kaggle or a GPU machine:
uv run python scripts/train_dpo.py

Key Hyperparameters (optimized for T4 GPU):

• beta: 0.1 (DPO temperature)
• learning_rate: 5e-6
• per_device_train_batch_size: 1
• gradient_accumulation_steps: 8
• max_length: 768

---

🤖 Inference

Download the trained adapter from Hugging Face and load it locally.

# Make sure the adapter is in models/llama3_dpo_adapter/
uv run python inference/inference.py

Or use it directly in your own code:

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "Karan6124/llama3-8b-dpo-orca-adapter",
    max_seq_length = 2048,
    load_in_4bit = True,
)
FastLanguageModel.for_inference(model)

messages = [{"role": "user", "content": "Why use DPO over SFT?"}]
inputs = tokenizer.apply_chat_template(
    messages, tokenize=True, add_generation_prompt=True, return_tensors="pt"
).to("cuda")

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.batch_decode(outputs, skip_special_tokens=True)[0])

---

📊 Benchmarking

Run the benchmark script on Kaggle to generate a comparison report:

uv run python evaluation/benchmark.py

Results are saved automatically to evaluation/benchmark_report.txt.

---

🏆 Key Lessons Learned

---

📄 License

This project is licensed under the MIT License — see LICENSE for details.

---

🙏 Acknowledgements

• Unsloth for making LLM fine-tuning incredibly fast
• TRL by HuggingFace for the DPO implementation
• Intel/orca_dpo_pairs for the training dataset
• Kaggle for the free GPUs that made this possible