Track hyperparameters, metrics, and model checkpoints across experiments with automated runs and artifact versioning. Automate hyperparameter optimization through sweeps, share findings via interactive reports, and manage production models through a centralized registry.
Weights & Biases (W&B) is a cloud-based MLOps platform — a suite of tools designed to bring order, reproducibility, and collaboration to the machine learning development lifecycle. If you have ever lost track of which model configuration produced your best results, accidentally overwritten a trained model, or struggled to share experiment results with a colleague, W&B was built to solve exactly those problems.
The name "Weights & Biases" is a nod to the fundamental learnable parameters of a neural network, but the platform serves any machine learning workflow, not just deep learning. At its core, W&B acts as a persistent, searchable record of everything that happens during ML development: what you tried, what worked, what didn't, and what artefacts were produced along the way.
Training a machine learning model is rarely a one-shot process. In practice, a researcher might run dozens or hundreds of experiments, varying:
Without a systematic way to track these variables and their outcomes, progress becomes difficult to reproduce and nearly impossible to share reliably. MLOps (Machine Learning Operations) is the discipline of applying engineering rigour to this process — and W&B is one of the leading platforms providing the tooling to do so.
W&B is organised around several distinct but interconnected products. Understanding what each one does — and how they relate — gives you a map of the platform before diving into any individual feature.
The five major components of W&B and how they radiate from a central platform hub. Each component addresses a distinct stage of the ML workflow.
A run is the fundamental unit in W&B — it represents a single execution of your training script. Every time you run an experiment, W&B can automatically capture:
All runs are stored in a project, where you can compare them side-by-side in an interactive dashboard. This eliminates the need for manual spreadsheets or ad-hoc logging scripts.
An artifact in W&B is a versioned, tracked data object — typically a dataset, a model checkpoint, or any file your pipeline produces or consumes. Artifacts solve a critical reproducibility problem: if you retrain a model six months later, you need to know exactly which version of the data was used and which preprocessing steps were applied.
W&B Artifacts provide:
Finding the right hyperparameters is an optimisation problem in its own right. W&B Sweeps automate this search by coordinating multiple runs across a defined search space. You specify the parameters to vary and their ranges, and W&B launches and monitors the search using strategies such as:
A sweep agent can run on a single machine or be distributed across many machines, all reporting back to the same W&B project.
Reports are interactive, narrative documents that live inside W&B. They allow you to combine live experiment charts, images, code snippets, and markdown prose into a single shareable document. Reports are useful for:
Unlike a static PDF or slide deck, W&B reports stay connected to your live data — charts can be re-rendered as new runs are added.
The Model Registry is a centralised catalogue for production-ready models. Once a model has been trained, evaluated, and approved, it can be registered with a semantic version and promoted through stages (e.g., staging → production). This bridges the gap between experimentation and deployment by giving teams a governed, single source of truth for which model is in use and why.
A typical W&B workflow might look like this:
v1).v1).Each step links to the others, creating a traceable chain from raw data to deployed model.
| Term | Meaning |
|---|---|
| Project | A collection of related runs, artifacts, and reports |
| Run | One execution of a training or evaluation script |
| Entity | A W&B user or team that owns projects |
| Artifact | A versioned, logged file (dataset, model, etc.) |
| Sweep | An automated hyperparameter search across multiple runs |
| Report | A shareable, interactive document combining prose and live charts |
W&B integrates with your existing code through a lightweight Python SDK. At its simplest, you add a few lines to your training script:
import wandb
wandb.init(project="my-project", config={"lr": 0.001, "epochs": 10})
for epoch in range(config.epochs):
loss = train(...)
wandb.log({"loss": loss, "epoch": epoch})
wandb.finish()
This minimal instrumentation is enough to capture a fully searchable, visualisable run. From there, each additional W&B feature builds on this foundation, adding versioning, automation, and collaboration capabilities progressively as your needs grow.
Weights & Biases brings software engineering best practices — version control, automated testing, centralised logging — into the machine learning workflow. Its five core components (Runs, Artifacts, Sweeps, Reports, and the Model Registry) each address a distinct pain point, and together they form a cohesive platform that scales from individual experimentation to team-level ML operations. The following concept documents explore each component in detail.