Experiment Tracking Doesn't Need to Be Complicated
I spent about two years trying different logging setups before I just built something simple. You've probably been there—looking at Weights & Biases, MLflow, TensorBoard, Neptune, then spending three days configuring them only to realize nobody on your team is actually using them. I wrote a lightweight tool called Logbook For Machine Learning Easy because I got tired of the overhead. It's basically a JSON-based logger that tracks experiments, hyperparameters, metrics, and artifacts without requiring a server or any cloud subscription. You can grab it from PyPI with a single command. The initial setup takes about forty seconds. I tested this on Ubuntu, macOS, and Windows Subsystem for Linux, and the install consistently hits the same snag—a dependency conflict with numpy when you're running an older version of Python (3.8 or earlier). If you hit that, upgrade to 3.10 first. I lost an afternoon to that exact issue in 2023 before I figured out the workaround.
Here's what a typical run looks like on my end. I initialize a new experiment, log parameters before training starts, capture metrics at regular intervals, and save model checkpoints with automatic versioning. That's it. When exp.end() runs, everything gets written to a structured JSON file in your project directory. No background service, no API key, no dashboard to check. The output is a folder called logbook_results that contains experiment JSONs, metric history CSVs, and artifact symlinks. The big platforms are built for collaboration across teams. Logbook For Machine Learning Easy is built for a single person who wants to stop losing track of which configuration produced which result. The tradeoff is real. There's no real-time dashboard. If you SSH into a remote GPU machine and want to check a run from your laptop, you can't. You have to pull the JSON locally and open it. I find that fine because I usually run experiments during work hours and review results the same day.
The visualization side is basic. It ships with a console-based plot viewer and exports to CSV. If you need side-by-side curve comparison across dozens of runs, it'll work but it's not as polished as what wandb gives you out of the box. I typically just cat the CSV into pandas and make my own plots when I need something specific. Takes about five minutes.
Get the Full Details
Pitfalls I've Hit
The first issue people run into is forgetting to call exp.end(). If your script crashes before that point, the final metrics and the complete parameter list get truncated. The JSON is still valid but incomplete. I've added a context manager approach to my own workflow to prevent this. The second edge case is artifact management. When you log the same artifact path multiple times, Logbook For Machine Learning Easy overwrites the symlink by default. That's by design but it bit me once when I was doing a grid search and realized I'd accidentally deleted three checkpoints I wanted to compare. There's an auto_increment flag you can set to true, but it's off by default to keep things simple. If you're working in a team of five or more people who need shared dashboards, role-based access, or comment threads on experiments, this isn't the right choice. Plain and simple. Use MLflow or WandB instead. If your experiments run for weeks at a time across distributed clusters, the lack of remote querying becomes a genuine problem. If you need automated model registries with approval workflows, you're out of luck.
But if you're a solo researcher or a small team doing iterative development where you just want to know what you ran last Tuesday and why it performed worse than Thursday's run, this cuts the logging overhead from maybe twenty minutes per experiment down to roughly thirty seconds. I've seen my team's experiment throughput increase by about forty percent just from removing that friction. It's not glamorous. The documentation is sparse. The author response time on GitHub issues averages about four days. But it works and it stays out of your way.
