FineTuna

Fine-tune any open-source LLM without GPU ownership, setup headaches, or ML expertise.

---

🎯 The Problem

Fine-tuning LLMs today requires:

• Expensive GPUs ($5k-50k upfront) 💸
• ML expertise (hyperparameters, techniques, optimization) 🧠
• Fragmented tools (notebooks, CLIs, dashboards scattered everywhere) 🔧
• Manual boilerplate (script generation, artifact management, data packaging) 📝 Most people never even try because the friction is too high.

---

✨ The Solution

FineTuna is a single, guided dashboard that handles the entire fine-tuning workflow:

Model Selection → Data Setup → Training Config → Kaggle Execution → Download Results

All in your browser. No GPU needed. No setup required. Free forever.

---

🚀 Features

🧭 Guided Workflow

• Step-by-step flow prevents confusion
• Clear progress indication
• One-click jumps between sections
• Mobile-responsive design

🔎 Model Discovery

• Search 10,000+ models from Hugging Face
• Filter by:
  • Model size (7B, 13B, 70B, etc.)
  • Architecture (Llama, Mistral, Phi, etc.)
  • Estimated VRAM requirements
  • Community ratings
  • Trending/Latest filters

🗂️ Intelligent Data Management

• Pre-built templates for 5 domains:
  • ⚖️ Legal (contracts, case law, legal reasoning)
  • 🏥 Medical (diagnoses, treatment plans, medical Q&A)
  • 💰 Finance (tax planning, financial analysis)
  • 💻 Code (debugging, code review, implementation)
  • 📝 General (customer support, writing)
• Custom data uploads (CSV, JSON, PDF)
• Template editor with live preview
• Community templates (created and rated by users)

⚙️ Advanced Configuration

• Fine-tuning techniques:
  • LoRA (balanced, recommended)
  • QLoRA (fastest, lowest memory)
  • Full fine-tuning (best quality)
• Auto-tuned hyperparameters with expert suggestions
• Gradient checkpointing and mixed precision
• Batch size calculator based on GPU type

☁️ Kaggle GPU Orchestration

• Free GPU hours leverage (30-37/month per user)
• One-click job submission
• Real-time training progress tracking
• Automatic error recovery
• Job history & rerun capability

📊 Performance Tracking

• Before/after loss comparison
• Training metrics dashboard
• Inference latency estimates
• Auto-generated technique reports
• Exportable results (PDF + JSON)

🎛️ Accelerator Selection

• GPU (T4, V100, P100)
• TPU (experimental)
• CPU (for testing)
• Explicit Kaggle accelerator mapping

📦 Smart Artifact Management

• Only download what you need
• Adapter files (LoRA: ~50-100MB)
• Merged models (full: 7-10GB)
• GGUF quantized versions (~2-3GB)
• Inference code snippets (Python, Node.js, GGUF)

🔐 Multi-Provider Authentication

• Supabase (database + auth)
• Kaggle (GPU execution)
• Google Drive (optional cloud storage)
• Hugging Face (model discovery)
• Resend (email notifications)

---

📊 Product Preview

---

🏃 Quick Start

Prerequisites

• Node.js 16+
• npm or yarn
• Kaggle account (free)
• Hugging Face account (optional)

Installation

# Clone the repository
git clone https://github.com/yourusername/finetuna.git
cd finetuna
 
# Install dependencies
npm install
 
# Set up environment variables
cp .env.example .env.local
 
# Start development server
npm run dev

Environment Variables

# Supabase
VITE_SUPABASE_URL=https://your-project.supabase.co
VITE_SUPABASE_ANON_KEY=your_anon_key
 
# Kaggle
VITE_KAGGLE_USERNAME=your_username
VITE_KAGGLE_KEY=your_api_key
 
# HuggingFace
VITE_HF_API_KEY=your_hf_token
 
# Google Drive (optional)
VITE_GOOGLE_CLIENT_ID=your_client_id
 
# Resend (optional)
VITE_RESEND_API_KEY=your_resend_key
 
# API
VITE_API_URL=http://localhost:8888

---

🌐 Deployment

Deploy to Netlify (Recommended)

# Install Netlify CLI
npm install -g netlify-cli
 
# Build
npm run build
 
# Deploy
netlify deploy --prod

Environment variables should be set in Netlify UI under Site settings > Build & deploy > Environment.

Deploy to Vercel

vercel --prod

---

🏗️ Architecture

┌─────────────────────────────────────────────────┐
│              FineTuna Frontend                   │
│         (React + Vite + TailwindCSS)            │
│                                                 │
│  • Model Discovery      • Training Dashboard    │
│  • Data Management      • Results Download      │
│  • Configuration        • Run History          │
└────────────────────┬────────────────────────────┘
                     │
        ┌────────────┼────────────┐
        │            │            │
    ┌───▼───┐   ┌───▼────┐  ┌───▼─────┐
    │Netlify│   │Supabase│  │ Kaggle  │
    │ API   │   │  Auth  │  │  GPU    │
    └───────┘   │ Storage│  │Execution│
                └────────┘  └─────────┘
                
    • Route handlers    • User data      • Kernel execution
    • Auth middleware   • Job metadata   • Model training
    • Job orchestration • Credentials    • Output storage

Technology Stack

---

📁 Project Structure

finetuna/
├── src/
│   ├── components/          # React components
│   │   ├── ModelSelector/   # Model discovery UI
│   │   ├── DataManager/     # Dataset upload & templates
│   │   ├── ConfigPanel/     # Training configuration
│   │   ├── ProgressDash/    # Real-time training status
│   │   └── ResultsView/     # Download & metrics
│   ├── pages/               # Page components
│   ├── hooks/               # Custom React hooks
│   ├── utils/               # Helper functions
│   └── App.tsx              # Main app
├── netlify/functions/
│   ├── api/auth/            # Authentication endpoints
│   ├── api/models/          # Model discovery
│   ├── api/jobs/            # Job management
│   ├── api/kaggle/          # Kaggle orchestration
│   ├── api/artifacts/       # Download handling
│   └── api/drive/           # Google Drive integration
├── docs/
│   ├── ARCHITECTURE.md      # System design details
│   ├── API.md               # API documentation
│   └── CONTRIBUTING.md      # Contribution guidelines
├── .env.example             # Environment template
├── netlify.toml             # Netlify configuration
└── vite.config.ts           # Vite configuration

---

🔐 Security & Privacy

✅ Environment files are never committed
✅ API keys stored securely server-side
✅ User data encrypted at rest
✅ Kaggle credentials never exposed to frontend
✅ Large artifacts only downloaded on explicit request
✅ No tracking (privacy-first)

See SECURITY.md for detailed security practices.

---

📚 Documentation

• API Reference - Complete endpoint documentation
• Architecture Guide - System design & data flow
• Contributing Guide - How to contribute
• Deployment Guide - Deploy your own instance
• FAQ - Common questions & answers

---

🎯 Use Cases

👨‍💻 Individual Developers

Fine-tune models for personal projects without buying GPUs.

🎓 Students

Learn LLM fine-tuning without expensive hardware or setup.

🏢 Teams

Create domain-specific models (legal assistant, medical QA, etc.).

🚀 Startups

Build AI features without ML infrastructure investment.

📚 Researchers

Quick experimentation with different techniques and datasets.

---

🤝 Contributing

We welcome contributions! Please see CONTRIBUTING.md for guidelines.

Quick Contribution Steps

# 1. Fork the repo
git clone https://github.com/YOUR_USERNAME/finetuna.git
 
# 2. Create feature branch
git checkout -b feature/amazing-feature
 
# 3. Make your changes & test
npm run dev
 
# 4. Commit & push
git commit -m "feat: add amazing feature"
git push origin feature/amazing-feature
 
# 5. Open pull request

Areas to Contribute

• 🐛 Bug fixes
• ✨ UI/UX improvements
• 📚 Documentation
• 🧪 Tests
• 🌍 Internationalization
• 💡 Feature ideas

---

📊 Stats

• ⭐ Stars: 0 (help us grow!)
• 🍴 Forks: 0
• 👥 Contributors: 1
• 📦 Package Size: ~500KB (gzipped)
• ⚡ Lighthouse Score: 95+
• 🌍 Supported Models: 10,000+

---

🐛 Troubleshooting

Kaggle GPU Not Available?

• Check your Kaggle free tier GPU hours
• Try switching to T4 (most stable)
• Wait during peak hours (try off-peak)

Model Download Too Large?

• Download adapter files instead of merged model
• Use GGUF quantized version
• Split downloads into parts

Training Loss Not Decreasing?

• Check dataset quality
• Reduce learning rate
• Increase training epochs
• Try QLoRA if OOM error See FAQ.md for more solutions.

---

📄 License

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

---

🙌 Acknowledgments

• Hugging Face - For the amazing model hub
• Kaggle - For free GPU resources
• Supabase - For auth and storage
• The open-source community - For amazing tools and libraries

---

📞 Get in Touch

• 🐦 Twitter: @ThaGeekiestOne
• 📧 Email: ayushnagarkoti2005@gmail.com
• 💡 Issues: GitHub Issues
• 💬 Discussions: GitHub Discussions

---

🌟 If FineTuna Helped You

Give us a star! ⭐ It means the world to our tiny team.

If FineTuna saved you time or money, consider:
• ⭐ Star this repo
• 🐦 Share on Twitter
• 📢 Recommend to a friend
• 🚀 Deploy your own instance

---