Skip to content

Dalva Logo

Dalva

A lightweight, self-hosted experiment tracker for deep learning

Dalva provides a simple Python API for logging experiments and a web interface for visualizing results.

Quick Start

import dalva

# Initialize a run (server must be running)
run = dalva.init(
    project="my-project",
    name="experiment-1",
    config={"lr": 0.001},
    server_url="http://localhost:8000"
)

# Log metrics
run.log({"loss": 0.5, "accuracy": 0.8}, step=0)
run.log({"loss": 0.3, "accuracy": 0.9}, step=1)

# Finish the run
run.finish()

Features

  • Simple API - Just init(), log(), and finish()
  • Self-Hosted - Data stored locally in DuckDB
  • Flexible Metrics - Log any metrics without schemas
  • Web Interface - View and compare experiments at http://localhost:8000

Installation

git clone https://github.com/tuliosouza99/Dalva.git
cd Dalva/backend && uv sync

Start the Server

dalva server start

Open http://localhost:8000 to access the web interface.

Important

The Dalva server must be running before logging experiments. The Python SDK communicates with the server via HTTP at the specified server_url.