Spaces:
Sleeping
Sleeping
| # 🚀 VISH AI - Quick Start Guide | |
| ## Installation & Setup (5 minutes) | |
| ### Step 1: Install Dependencies | |
| ```bash | |
| pip install -r requirements.txt | |
| ``` | |
| **What gets installed:** | |
| - Gradio (UI) | |
| - FastAPI (API) | |
| - Transformers (Phi-3) | |
| - PEFT (LoRA training) | |
| - Datasets (data handling) | |
| ### Step 2: Start the Server | |
| ```bash | |
| python start.py | |
| ``` | |
| **Or manually:** | |
| ```bash | |
| python -m app.main | |
| ``` | |
| ### Step 3: Open Browser | |
| Visit: **http://localhost:7860** | |
| --- | |
| ## 🎯 First Steps | |
| ### 1. Try the Chat | |
| - Go to "💬 VISH Assistant" tab | |
| - Type: "Tell me about artificial intelligence" | |
| - Click Send | |
| - Notice the interaction ID in the response | |
| ### 2. Submit Feedback | |
| - Copy the interaction ID (e.g., `a1b2c3d4`) | |
| - Go to "⭐ Feedback" tab | |
| - Paste the ID | |
| - Rate 1-5 stars | |
| - Click "Submit Feedback" | |
| ### 3. Check Statistics | |
| - Go to "📊 Statistics" tab | |
| - Click "🔄 Refresh Stats" | |
| - See your interactions and ratings | |
| ### 4. Train the Model (After 10+ interactions) | |
| - Go to "🎓 Training (Admin)" tab | |
| - Set minimum samples: 10 | |
| - Set epochs: 3 | |
| - Enter admin key: `vish-admin-2024` (default) | |
| - Click "🚀 Start Training" | |
| - Wait 10-30 minutes for training | |
| --- | |
| ## 📝 Category Examples | |
| ### General Assistant | |
| ``` | |
| Category: assistant | |
| Question: "What is machine learning?" | |
| ``` | |
| ### Resume Builder | |
| ``` | |
| Category: resume | |
| Question: "Help me write a software engineer resume" | |
| ``` | |
| ### Research | |
| ``` | |
| Category: research | |
| Question: "Explain quantum computing" | |
| ``` | |
| ### Business | |
| ``` | |
| Category: business | |
| Question: "How do I create a business plan?" | |
| ``` | |
| --- | |
| ## 🔐 Admin Key | |
| Default admin key: `vish-admin-2024` | |
| **Change it:** | |
| ```bash | |
| export VISH_ADMIN_KEY="your-secret-key" | |
| ``` | |
| Or in `.env` file: | |
| ``` | |
| VISH_ADMIN_KEY=your-secret-key | |
| ``` | |
| --- | |
| ## 📊 Understanding the System | |
| ### Data Flow | |
| 1. **User chats** → Saved to `data/vish_dataset.jsonl` | |
| 2. **User rates** → Saved to `data/feedback.jsonl` | |
| 3. **Training runs** → Creates `models/vish-ai-mini/latest/` | |
| 4. **Model reloads** → Uses improved version automatically | |
| ### File Structure | |
| ``` | |
| data/ | |
| ├── vish_dataset.jsonl # All interactions | |
| ├── feedback.jsonl # User ratings | |
| └── research_data.jsonl # Research data | |
| models/ | |
| └── vish-ai-mini/ | |
| ├── latest/ # LoRA adapters | |
| └── metadata.json # Version info | |
| ``` | |
| --- | |
| ## 🎓 Training Process | |
| ### When to Train | |
| - After collecting 10+ interactions | |
| - After significant feedback | |
| - Weekly/monthly for continuous improvement | |
| ### Training Time | |
| - **CPU**: 10-30 minutes | |
| - **GPU**: 2-5 minutes | |
| ### What Gets Trained | |
| - High-quality interactions (rating ≥ 3) | |
| - Deduplicated data | |
| - LoRA adapters only (efficient!) | |
| ### Model Versions | |
| Each training creates a version: | |
| - `v20241016_143022` | |
| - `v20241017_095234` | |
| - Latest version is used automatically | |
| --- | |
| ## 🚀 Deployment | |
| ### Hugging Face Spaces | |
| 1. Create Space: https://huggingface.co/new-space | |
| 2. Upload files: | |
| - `app/` folder | |
| - `requirements.txt` | |
| - `Dockerfile` | |
| - `README.md` | |
| 3. Set hardware: CPU Basic (free) or T4 GPU | |
| 4. Wait for build (~15-20 minutes first time) | |
| 5. Done! Your AI is live | |
| ### Docker | |
| ```bash | |
| # Build | |
| docker build -t vish-ai . | |
| # Run | |
| docker run -p 7860:7860 \ | |
| -v $(pwd)/data:/app/data \ | |
| -v $(pwd)/models:/app/models \ | |
| -e VISH_ADMIN_KEY=your-key \ | |
| vish-ai | |
| ``` | |
| --- | |
| ## ⚡ Quick Tips | |
| 1. **Start with general questions** to build dataset | |
| 2. **Rate honestly** - only good data improves the model | |
| 3. **Train regularly** - weekly is good | |
| 4. **Check stats** - monitor improvement | |
| 5. **Backup data** - copy `/data` and `/models` regularly | |
| --- | |
| ## 🐛 Common Issues | |
| ### "Model not loaded" | |
| - Wait for initial download (~7GB, 10-15 min) | |
| - Check logs for errors | |
| - Verify internet connection | |
| ### "Insufficient data for training" | |
| - Need at least 10 interactions | |
| - Check: `curl http://localhost:7860/api/stats` | |
| ### "Out of memory" | |
| - Use quantization (edit `model_handler.py`) | |
| - Reduce batch size in `retrain.py` | |
| - Upgrade to GPU | |
| --- | |
| ## 📚 Next Steps | |
| 1. **Explore API**: Visit `http://localhost:7860/docs` | |
| 2. **Read Full README**: See `README_SELF_TRAINING.md` | |
| 3. **Customize**: Edit system prompts in `gradio_ui.py` | |
| 4. **Integrate**: Use API endpoints in your apps | |
| --- | |
| ## 🎉 Success! | |
| You now have a self-improving AI assistant that: | |
| - ✅ Learns from your conversations | |
| - ✅ Improves with your feedback | |
| - ✅ Trains automatically with LoRA | |
| - ✅ Tracks performance over time | |
| - ✅ Works on free-tier hardware | |
| **Happy chatting! 🤖** | |
| --- | |
| Built with ❤️ by Vishwas | Questions? Open an issue! | |