Model Training Manager

A professional-grade application for managing AI model training with worker orchestration, dataset management, and seamless Ollama integration. Supports QLoRA, Unsloth, RAG, and standard fine-tuning workflows through a clean React/TypeScript UI.

Features

• Dataset management — upload and manage CSV/JSON training datasets
• Training job orchestration — create, monitor, and manage training jobs
• Worker pool management — scale training with concurrent worker threads
• Multiple training types — QLoRA, Unsloth, RAG, and standard fine-tuning
• Real-time monitoring — track training progress, epochs, and loss metrics
• Interactive API docs — Swagger UI and ReDoc included

Tech Stack

Layer	Technology
Backend	Python 3.11+ / FastAPI / SQLAlchemy (SQLite) / Pydantic
Frontend	React 18+ / TypeScript / Vite / TailwindCSS / React Query
LLM Backend	Ollama (local)

Requirements

• Python 3.11+
• Node.js 20+ and npm
• Ollama running locally
• 8GB+ RAM recommended

---

Quick Start

1. Install and start Ollama

curl -fsSL https://ollama.ai/install.sh | sh
ollama serve
ollama pull llama3.2:3b   # in a new terminal

2. Backend

git clone <repository-url>
cd trainers/backend
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt
uvicorn app.main:app --reload --host 0.0.0.0 --port 8000

3. Frontend

cd ../frontend
npm install
npm run dev

Access:

• Frontend: http://localhost:5173
• API docs: http://localhost:8000/api/docs
• ReDoc: http://localhost:8000/api/redoc

---

Configuration

Create a .env file in backend/:

Variable	Default	Description
DATABASE_URL	sqlite:///./trainers.db	Database connection
UPLOAD_DIR	./uploads	Dataset upload directory
OLLAMA_BASE_URL	http://localhost:11434	Ollama API URL
OLLAMA_MODEL	llama3.2:3b	Default training model
MAX_WORKERS	8	Maximum concurrent workers
DEBUG	false	Enable debug mode

Training parameters (configurable via Settings page):

Parameter	Range	Default
Batch Size	1–64	4
Learning Rate	1e-6 to 1.0	2e-4
Epochs	1–100	3
LoRA Rank	1–256	16
LoRA Alpha	1–512	32
LoRA Dropout	0.0–1.0	0.05

---

API Reference

Method	Endpoint	Description
GET	/health	Health check
POST	/api/v1/datasets/	Upload dataset
GET	/api/v1/datasets/	List datasets
POST	/api/v1/jobs/	Create training job
GET	/api/v1/jobs/	List training jobs
GET	/api/v1/jobs/{id}/status	Get job status
POST	/api/v1/jobs/{id}/start	Start job manually
POST	/api/v1/jobs/{id}/cancel	Cancel job
GET	/api/v1/config/	Get configuration
PATCH	/api/v1/config/	Update configuration
GET	/api/v1/workers/	Get worker status
POST	/api/v1/workers/	Control workers

---

Training Types

QLoRA — 4-bit quantization with low-rank adaptation. Best for limited GPU memory scenarios.

Unsloth — Optimized LoRA training with faster throughput and reduced memory usage. Best for efficient fine-tuning.

RAG — Builds a vector index from the dataset and trains retrieval and generation components. Best for knowledge-intensive tasks.

Standard fine-tuning — Traditional supervised fine-tuning with rule-based configuration. Best for straightforward adaptation tasks.

---

Testing

# Backend
cd backend
source venv/bin/activate
pytest --cov=app --cov-report=term-missing

# Frontend
cd frontend
npm test

---

Project Structure

trainers/
├── backend/
│   ├── app/
│   │   ├── api/           # API endpoints
│   │   ├── core/          # Configuration, database
│   │   ├── models/        # SQLAlchemy models
│   │   ├── schemas/       # Pydantic schemas
│   │   ├── services/      # Business logic
│   │   └── workers/       # Training workers
│   ├── tests/
│   └── requirements.txt
└── frontend/
    ├── src/
    │   ├── components/    # React components
    │   ├── pages/         # Page components
    │   ├── services/      # API client
    │   └── types/         # TypeScript types
    └── package.json

---

Troubleshooting

Ollama connection refused:

ollama serve
curl http://localhost:11434/api/tags

Database errors:

rm backend/trainers.db   # reset — tables recreate on next start

Port conflicts:

lsof -i :5173   # frontend
lsof -i :8000   # backend
lsof -i :11434  # ollama

Out of memory: Reduce batch size, reduce max workers, or switch to QLoRA training type.

Python import errors: Ensure the virtual environment is activated before running.

Frontend build errors:

cd frontend && rm -rf node_modules package-lock.json && npm install

---

License

MIT