🚀 SFT Template - Supervised Fine-Tuning with Unsloth

This project provides a professional, modular, and configurable template for Supervised Fine-Tuning (SFT) of Large Language Models (LLMs). It is specifically optimized to leverage Unsloth, enabling faster and more memory-efficient training on Consumer or Datacenter GPUs.

---

✨ Key Features

• Modular Design: Clean separation between Model handlers, Dataset processors, and Training loops.
• Config-Driven Workflow: Manage all hyperparameters—from Model quantization to PEFT/LoRA settings—via a single src/config.yaml file.
• Abstract Base Classes: Built on extensible ABC interfaces, allowing for easy integration of new dataset formats or custom training logic.
• Unsloth Optimization: Automated integration for 4-bit loading, LoRA/PEFT application, and memory-efficient training.
• Reproducibility: Each run is timestamped and automatically backs up its configuration and dataset for future reference.
• Scaling with Environment Variables: Override any training parameter (e.g., LEARNING_RATE, NUM_TRAIN_EPOCHS) on-the-fly via system environment variables.

---

📂 Project Structure

.
├── src/
│   ├── core/
│   │   └── base.py           # Abstract Base Classes (Standards Definition)
│   ├── model/
│   │   └── handler.py        # Model/Tokenizer initialization (Unsloth/PEFT)
│   ├── dataset/
│   │   └── processor.py      # Data loading, cleaning, and LoRA formatting
│   ├── train/
│   │   └── sft.py            # Training orchestration and persistence
│   ├── prompts/              # Centralized prompt templates
│   ├── config.yaml           # Centralized configuration file
│   └── main.py              # Application Entry Point
├── data/                    # Recommended directory for raw datasets
├── requirements.txt         # Project dependencies
└── README.md                # Documentation (You are here)

---

🛠️ Installation

Prerequisites

• Python 3.10+
• CUDA-compatible GPU

Setup

We recommend using a Virtual Environment (Conda or venv):

# Example using Conda
conda create -n sft_template python=3.10 -y
conda activate sft_template

# Install core dependencies
pip install unsloth trl transformers datasets torch pyyaml

---

⚙️ Configuration & Usage

1. Configure the Run

Edit src/config.yaml to suit your training needs. Key sections include:

• model: Choose your base model (e.g., Gemma, Llama 3) and sequence length.
• dataset: Provide the path to your JSONL file and define train/test splits.
• training: Standard HF/TRL training arguments.
• peft: LoRA rank (r), alpha, and dropout.

2. Execution

Run the training script via the CLI:

python src/main.py --config src/config.yaml

3. Dynamic Overrides (Advanced)

You can override any parameter in the training section by prefixing the name with its uppercase version in the environment:

# Override learning rate and epochs without changing the YAML
LEARNING_RATE=5.0e-6 NUM_TRAIN_EPOCHS=10 python src/main.py

---

🧩 Extending the Template

Custom Datasets

If your dataset format differs from the default JSONL structure, simply extend AbstractDatasetProcessor in src/dataset/processor.py:

class MyCustomProcessor(AbstractDatasetProcessor):
    def format_dataset(self, dataset):
        # Your custom mapping logic here
        ...

Custom Prompts

Add new templates to src/prompts/ to maintain a clean separation between prompting strategies and data logic.

---

📈 Monitoring

Logs are saved to the directory specified in output.output_dir (default: ./outputs/). By default, the project supports Tensorboard reporting.

---