Built a Web App to Train & Fine-Tune Ollama Models
How I built a full-stack web platform for training, fine-tuning, and managing Ollama AI models with real-time GPU monitoring and job tracking.
The Problem
Fine-tuning LLMs is complex — you need to manage datasets, configure training jobs, monitor GPU usage, track model versions, and deploy to Ollama. I was doing all of this manually through terminal commands and scripts.
I needed a proper dashboard to manage the entire workflow.
What I Built
A full-stack Ollama Model Training Platform — a web app where you can:
- Upload and manage training datasets
- Create and queue fine-tuning jobs (LoRA, QLoRA)
- Monitor training progress in real-time
- Track GPU server status and utilization
- Convert trained models to Ollama format
- Deploy models with one click
Dashboard Overview
The platform dashboard shows everything at a glance:
- 3 Datasets uploaded and ready for training
- 4 Training Jobs — active, completed, and failed
- 1 Trained Model ready for deployment
- 2 Active Jobs currently running
- 1/2 GPU Servers online
Platform Features
Training Job Management
- Queue multiple training jobs with different configurations
- Support for QLoRA and LoRA fine-tuning methods
- Real-time progress bars with ETA estimates
- Job status tracking: queued, running, completed, failed
GPU Monitoring
- Real-time GPU utilization and memory tracking
- Multi-server support for distributed training
- Automatic job scheduling based on GPU availability
Model Management
- Track all trained models with metadata
- One-click Ollama model conversion
- Version history and rollback support
Tech Stack
- Frontend: Next.js, TypeScript, Tailwind CSS, shadcn/ui
- Backend: Node.js, Express, PostgreSQL
- AI/ML: Ollama, Hugging Face Transformers, PyTorch
- Real-time: Pusher for live progress streaming
- GPU: Runpod integration for cloud GPU instances
Recent Jobs
The platform tracks all training jobs with their base models:
| Job | Base Model | Status |
|---|---|---|
| Support Bot v3 QLoRA | gemma2:2b | Queued |
| Code Assistant LoRA | qwen2.5:3b | Running (40%) |
| Large Model Experiment | llama3.1:8b | Failed |
| Chatbot Fine-tune v2 | gemma2:2b | Completed |
What I Learned
- Job queuing is essential — GPU resources are expensive, so smart scheduling matters
- Real-time monitoring saves hours of SSH-ing into servers to check progress
- LoRA/QLoRA makes fine-tuning accessible even on consumer GPUs
- Ollama conversion pipeline needs careful handling of model formats and quantization
This platform is now used internally at Glixen Tech for all our AI model training workflows.