Git Commit Message Generator

AI-powered CLI tool that automatically generates professional git commit messages from code diffs using a fine-tuned Qwen-0.5B language model.

Live Demo: HuggingFace Model

Features

• Accurate - Trained on 100K+ real commit messages
• Fast - Generates messages in ~2 seconds
• Consistent - Deterministic output (same diff = same message)
• Clean - Professional, concise commit messages
• Easy - Simple CLI: commit-gen generate

Demo

project_demo.mov

Click to watch the tool in action - generating commit messages from git diffs

Quick Start

Installation

# Clone the repository
git clone https://github.com/rajtiwariee/GitCommitGenerator
cd GitCommitGenerator

# Install dependencies
pip install -r requirements.txt

# Install CLI tool
pip install -e .

Usage

# Generate message for staged changes
git add file.py
commit-gen generate

# Generate and auto-commit
commit-gen generate --commit

# Interactive mode (edit before committing)
commit-gen generate --commit --interactive

# View configuration
commit-gen config --show

Model Training

Dataset Preparation

# Download CommitPackFT dataset
python scripts/download_dataset.py

# Clean and filter data
python scripts/prepare_dataset.py

# Format for training
python scripts/format_for_training.py

Fine-tuning

# Train with LoRA (recommended: ~7.5 hours on T4 GPU)
python train.py \
  --num_epochs 3 \
  --batch_size 4 \
  --gradient_accumulation_steps 8 \
  --use_lora \
  --gradient_checkpointing \
  --max_length 384 \
  --use_fp16

Evaluation

# Run automated metrics
python scripts/test_model.py --num_samples 100

# Interactive testing
python scripts/qualitative_test.py

# Plot training loss
python scripts/plot_loss.py

Model Details

Base Model: Qwen-0.5B
Training Method: LoRA (Low-Rank Adaptation)
Dataset: CommitPackFT (55K filtered samples)
Hardware: T4 GPU (16GB VRAM)
Training Time: ~7.5 hours
Parameters: 493M (base) + 35MB (LoRA adapters)

Training Configuration:

• Epochs: 3
• Batch Size: 4 (effective: 32 with gradient accumulation)
• Learning Rate: 5e-5
• LoRA r=16, alpha=32
• Mixed Precision (FP32 + FP16)

Training Loss

The model converged smoothly from an initial loss of 1.68 to a final loss of ~0.87 over 3 epochs.

Project Structure

GitCommitGenerator/
├── commit_gen/              # CLI tool package
│   ├── cli.py              # Main CLI interface
│   ├── generator.py        # Model inference
│   ├── git_utils.py        # Git integration
│   └── config.py           # Configuration
├── scripts/                # Utilities
│   ├── download_dataset.py
│   ├── prepare_dataset.py
│   ├── test_model.py
│   ├── qualitative_test.py
│   └── upload_to_huggingface.py
├── train.py                # Training script
├── setup.py                # Package setup
└── requirements.txt        # Dependencies

Documentation

• CLI_README.md - Detailed CLI usage guide
• DEMO_SCRIPT.md - Demo/presentation script

Publishing to HuggingFace

# Set your token
export HF_TOKEN="your_token_here"

# Upload model
python scripts/upload_to_huggingface.py \
  --model_path ./qwen-0.5b-finetuned/final \
  --repo_name your-username/model-name \
  --metrics_file evaluation_results.json

License

MIT

Author

Raj Tiwari (@rajtiwariee)