Wandb
Last Updated: July 29, 2026 | By Mihail Sebastian | AI Dictionary
Weights & Biases (W&B) is a platform for tracking machine learning experiments, visualizing training runs, and sharing results across a team.
What is Wandb?
Weights & Biases (W&B, often written “wandb” after its Python package) is a platform for tracking machine learning experiments: training code logs metrics, hyperparameters, and outputs as it runs, and W&B turns them into live dashboards a team can watch, compare, and share.
Its center of gravity is the experimentation phase – understanding which of many training runs worked and why, before anything ships.
How Weights & Biases Works
A few lines of code connect a training script to W&B. From then on, each run streams its hyperparameters, loss curves, and evaluation metrics to a project dashboard, where runs plot side by side.
On top of tracking, W&B adds Sweeps, which searches hyperparameter combinations across many runs, and Artifacts, which versions the datasets and models a run consumed and produced. Reports turn selected charts into shareable documents, which is where the collaboration strength shows.
Weights & Biases vs MLflow
The practical difference: W&B is first a hosted service, strongest at experiment dashboards and team collaboration, while MLflow is open-source software you run yourself that covers the whole lifecycle, including a built-in model registry. Teams that want rich visualization with no infrastructure to operate lean toward W&B; teams that want the system of record inside their own walls lean toward MLflow.
| Criterion | Weights & Biases | MLflow |
|---|---|---|
| Scope | Experiment tracking and visualization first | Full ML lifecycle, including a model registry |
| Delivery | Hosted platform by default | Open source; self-hosted by default |
| Known for | Dashboards and team collaboration | Running inside your own infrastructure |
Example of Weights & Biases
A research team runs fine-tuning jobs on a language model overnight across four GPU servers. Each job logs to the same W&B project, so by morning the dashboard shows every loss curve on one chart, with the run that diverged obvious at a glance.
One researcher marks the two best runs, writes a short W&B report comparing them, and sends the link. The discussion about which configuration to keep happens on the actual curves, not on screenshots pasted into chat.
Related AI terms: MLflow · Model Registry · Pipeline · Hyperparameter Tuning · Model Training
Did you like the Wandb gist?
Learn about 250+ need-to-know artificial intelligence terms in the AI Dictionary.
Mihail Sebastian — Writes about AI governance, regulation, and the technology behind them. Placeholder bio — replace with a real credential line. About