NuBoard

Experiment tracking that lives in a folder.
Log from Python with three calls. Every run is a folder. View thousands of runs in your browser.

import nuboard as nb

with nb.init("my-project", run_name="baseline", config={"lr": 3e-4}) as run:
    for step in range(10_000):
        loss = train_step()
        run.add_scalar({"train/loss": loss}, step=step)  # a few µs, never blocks training

nuboard up experiments        # http://localhost:6070

Why

• No setup. No account, no server, no database. pip install nuboard gives you logging and the viewer; logging uses only the standard library, so it adds nothing to the import time of a training script.
• Folder-contained. Each run is one folder (run.json, an append-only binary log, a media pack, a Parquet file at the end). Copy it, rsync it, zip it, delete it. Notes, tags, saved views and hypothesis boards live in the project folder too, so they can be committed to git.
• Fast. Logging is an in-memory append; a background thread writes columnar frames. The viewer reads a million points in about a tenth of a second and lists thousands of runs instantly (see the benchmarks).
• Records what you forget. Config, git commit and uncommitted diff, installed packages, GPUs, command line, rank, console output and CPU, memory and GPU use, automatically.
• Crash-safe. A killed process loses at most the last second of data; torn writes are ignored; crashes are marked with their traceback.

Quick start

pip install nuboard                   # or: uv pip install nuboard
nuboard demo demo-runs --up --open    # a demo workspace with a live run

Then try where tags has "sweep" group by config.optim.lr in the query bar and open the Charts page. The quickstart has a tour, and installation covers building from source.

Features

What changed in each release is in the changelog, and the known limitations page lists what doesn't work yet.

Documentation

The docs are a MkDocs site in docs/:

uv run mkdocs serve    # http://localhost:8000

• Installation and Quickstart
• Logging from Python, Using the viewer, Sweeps, Health alerts, Artifacts and lineage, Analysis API, Reports, Hypotheses, Reproducibility, Distributed training, Evaluating while training, Working on several machines, Importing from TensorBoard
• Python API, one page per call
• Reference: command line, query language, file format, architecture, benchmarks, development

Development

uv sync                                              # Python + dev tools
(cd frontend && pnpm install && pnpm build)          # builds the viewer into src/nuboard/static/
uv run pytest && uv run ruff check . && uv run pyright
(cd frontend && pnpm test && pnpm check && pnpm lint && pnpm e2e)
uv build                                             # wheel with the frontend bundled

See Development for the repository layout, the rules for the logging side, and how to rename the project (the name lives in three places).

License

NuBoard is released under the MIT License.

Requirements

Python 3.10+ on Linux, macOS or Windows. It installs NumPy, PyArrow, Starlette and uvicorn for the viewer; logging uses only the standard library. Building the frontend from source needs Node 20+ and pnpm (brew install pnpm on macOS); installing from PyPI does not.