GR Rebuild

A project for fine-tuning models with LoRAs using Unsloth, transformers, and Dstack.

Features

• Unsloth Integration: Efficient LoRA fine-tuning with Unsloth
• CLI Interface: Easy-to-use command-line tools
• Dstack Integration: Cloud compute with RunPod backend
• Slack-style Training: Optimized for conversational AI training
• Modular Design: Clean, extensible codebase

Installation

# Install uv (if not already installed)
curl -LsSf https://astral.sh/uv/install.sh | sh

# Install dependencies
uv sync

Usage

CLI Commands

# Check project status
grr status

# Train with sample data (for testing)
grr train --model unsloth/llama-3.2-1b-bnb-4bit --dataset sample

# Train with custom dataset
grr train --dataset custom --data-path your_data.json --epochs 5

# Create sample dataset file
grr create-sample --output my_data.json

# Evaluate a model (coming soon)
grr evaluate --model-path /path/to/model

Python API

from grr import UnslothFineTuner, TrainingConfig

# Create configuration
config = TrainingConfig(
    model_name="unsloth/llama-3.2-1b-bnb-4bit",
    max_seq_length=2048,
    num_train_epochs=3,
    output_dir="./output"
)

# Create fine-tuner
fine_tuner = UnslothFineTuner(config)

# Prepare training data
training_data = [
    {
        "instruction": "What is the capital of France?",
        "input": "",
        "output": "The capital of France is Paris."
    },
    # ... more examples
]

# Train the model
fine_tuner.train(training_data)

Dataset Format

Your training data should be in JSON format:

[
    {
        "instruction": "User question or instruction",
        "input": "Optional additional context",
        "output": "Expected model response"
    }
]

For evaluation data, you can provide separate train/eval sets:

{
    "train": [...],
    "eval": [...]
}

Development

# Run tests
uv run pytest

# Format code
uv run ruff format .

# Lint code
uv run ruff check .

# Auto-fix linting issues
uv run ruff check . --fix

# Build package
uv run hatch build

Project Structure

gr-rebuild/
├── src/
│   └── grr/
│       ├── __init__.py
│       ├── cli.py
│       └── finetune.py          # Unsloth fine-tuning module
├── examples/
│   └── finetune_example.py      # Example usage
├── tests/
│   ├── __init__.py
│   └── test_cli.py
├── pyproject.toml
└── README.md

Dstack Integration

To use with Dstack cloud compute:

1. Configure RunPod backend in ~/.dstack/server/config.yml
2. Start Dstack server

uv run dstack server

3. Start development environment in a new terminal:

   uv run dstack apply -f dev.dstack.yml

4. Connect via SSH or Cursor:

   ssh fudgin
   # or use the Cursor link

5. Install dependencies in the cloud environment:

   uv sync

6. Run training:

   grr train --dataset sample

Build Datasets

# Build dataset from Discord export (chat format)
uv run grr build-dataset \
  --source discord \
  --input data.json \
  --target-user duffdogg \
  --format chat \
  --output discord_chat.json

# Build dataset from Slack export (alpaca format)
uv run grr build-dataset \
  --source slack \
  --input slack_messages.jsonl \
  --target-user U06BOT \
  --format alpaca \
  --output slack_alpaca.json

Training Configuration

The TrainingConfig class supports various parameters:

• Model: unsloth/llama-3.2-1b-bnb-4bit (default)
• LoRA: r=16, alpha=32, dropout=0.1
• Training: 3 epochs, batch size 2, learning rate 2e-4
• Sequence Length: 2048 tokens

License

MIT License