Weights & Biases (often abbreviated as W&B or wandb) is a cloud-based platform for machine learning experiment tracking, visualization, and collaboration. Developed by Weights & Biases Inc., it enables practitioners to log hyperparameters, metrics, code versions, and model outputs in real time, and provides interactive dashboards for comparing runs, debugging models, and sharing results. The platform integrates with popular ML frameworks such as PyTorch, TensorFlow, and Keras, and offers additional tools for hyperparameter optimization (Sweeps), artifact versioning, and model management.
1.1 Founding and Company Background
Weights & Biases Inc. was founded in 2017 by Lukas Biewald, Chris Van Pelt, and Shawn Lewis. Biewald previously founded Figure Eight (formerly CrowdFlower), a data annotation platform, and brought experience in scaling human‑in‑the‑loop machine learning workflows. The company was established with the goal of creating a dedicated infrastructure layer for ML experiment management, addressing the lack of standardized tooling for tracking and reproducing deep learning experiments. The initial development was supported by seed funding from venture capital firms including Redpoint and Sapphire Ventures.
1.2 Key Milestones and Releases
The public beta of the platform launched in 2018, quickly gaining adoption in the deep learning community. In 2019, Weights & Biases introduced Sweeps, an automated hyperparameter optimization module. The Artifacts feature for dataset and model versioning followed in 2020. The Model Registry and Launch capabilities were rolled out in 2022, extending the platform beyond experiment tracking into model deployment and lifecycle management. By 2024, the platform reported over one million registered users and integration with more than fifty machine learning frameworks and libraries.
2.1 Experiment Tracking
Experiment tracking is the foundational capability of Weights & Biases. Users instrument their training scripts with a few lines of code to log metrics, hyperparameters, code state, and system resources in real time. Each training run is recorded as a distinct run with a unique ID, timestamps, and associated metadata. Runs can be organized by project, and all data is persisted on the cloud for later analysis.
2.1.1 Logging Metrics and Hyperparameters
Metrics (such as loss, accuracy, or learning rate) are logged using wandb.log() and can be scalar values, histograms, or media objects (images, audio, video). Hyperparameters are defined in a wandb.config object and automatically recorded. The platform supports real-time updates, enabling live visualization of training progress. Logged data is time‑stamped and can be filtered, grouped, and compared across runs.
2.1.2 Artifacts and Data Versioning
The Artifacts system provides version control for datasets, models, and other files. An artifact is a named, versioned collection of files or references to external storage (e.g., S3). Users can log artifacts as outputs of a run or use them as inputs, enabling full provenance tracking. Each artifact version is immutable and includes metadata such as creation time, checksum, and the run that produced it.
2.2 Visualization and Dashboards
The web interface offers interactive dashboards that render logged metrics and media in real time. Charts automatically update as new data arrives, and users can customize the layout, zoom, and axis ranges. All visualizations are shareable via persistent URLs.
2.2.1 Run Comparison and Parallel Coordinates
Multiple runs can be superimposed on line charts, scatter plots, or bar plots. The Parallel Coordinates plot is a particularly powerful tool for hyperparameter analysis: it maps each hyperparameter to a vertical axis and connects runs as lines, allowing quick identification of parameter combinations that correlate with high performance. The dashboard also supports grouping runs by tags, config values, or run groups.
2.2.2 Custom Charts and Plots
Users can create custom visualizations using the Panels framework. Panels are Python‑based chart objects (e.g., scatter, histogram, confusion matrix) that can be added to a project dashboard. The wandb.plot interface exposes pre‑built chart types, and advanced users can write custom panel scripts using the Weights & Biases API.
2.3 Collaboration and Reporting
The platform is built around team collaboration. All logged data lives in shared projects, and team members can view, comment on, and download results. Permissions are role‑based (admin, member, viewer).
2.3.1 Team Workspaces and Projects
Workspaces allow organizations to group related projects under a single team account. Each project contains all runs, artifacts, and dashboards for a particular research effort or model family. Team members can be assigned to projects with specific access levels. The Workspace view aggregates dashboards from multiple projects for cross‑project analysis.
2.3.2 Automated Reports and Notes
Reports are living documents that capture a snapshot of charts, tables, and markdown notes. Reports can be created manually or automatically generated from run comparisons. They support rich text, inline images, and embedded media. Reports are versioned and can be shared via link or exported to PDF. Users can also add inline notes to individual runs or reports for contextual commentary.
3.1 Supported Machine Learning Frameworks
Weights & Biases integrates directly with the most widely used ML frameworks through dedicated wandb callback classes or wrappers. Official integrations include PyTorch (via wandb.watch and WandbCallback), TensorFlow/Keras (via WandbCallback), JAX/Flax, Hugging Face Transformers, Fast.ai, XGBoost, LightGBM, and scikit‑learn. Integration typically requires only a few lines of additional code, and the platform automatically captures framework‑specific metrics (e.g., gradient histograms in PyTorch).
3.2 Python SDK and API
The primary interface to Weights & Biases is the Python software development kit (SDK), distributed as the wandb package. The SDK provides a high‑level API for logging, configuration, and orchestration. An HTTP‑based REST API is also available for programmatic access to logged data, user management, and custom integrations. The SDK supports asynchronous logging to minimize overhead in training loops.
3.3 Sweeps: Hyperparameter Optimization
Sweeps is an integrated hyperparameter search engine. Users define a configuration specifying the search space (continuous, discrete, or categorical parameters), the optimization algorithm (grid, random, Bayesian, or evolutionary), and the metric to optimize. When a sweep is launched, the platform distributes trials across one or more agents (local or remote), logs each trial’s metrics, and updates the search strategy in real time based on results. Sweep progress is visualized on a dedicated dashboard showing parameter importance and best‑performing runs.
3.4 Launch: Model Orchestration and Deployment
Launch extends the platform into model execution beyond experiment tracking. It enables users to define and run training or inference jobs on various compute backends (local machine, Kubernetes, or cloud clusters such as AWS, GCP, and Azure). Launch supports reproducible execution by packaging code and dependencies into a containerized environment. Jobs can be triggered manually, scheduled, or integrated into CI/CD pipelines.
3.5 Model Registry and Model Management
The Model Registry provides a centralized catalog for versioning, annotating, and promoting trained models. Each model version is linked to the run and artifacts that produced it, ensuring traceability. Models can be assigned lifecycle stages (e.g., staging, production, archived). The registry integrates with the Launch module to deploy specific model versions to production endpoints.
4.1 Initializing a Project and API Key
A user begins by signing up for an account on the Weights & Biases website and obtaining an API key. The key is stored locally using wandb login or an environment variable. A new project is created either through the web interface or automatically at the start of the first run by calling wandb.init(project="project_name").
4.2 Instrumenting Training Code
Within the training script, the user imports wandb and calls wandb.init() at the start of the experiment. Hyperparameters are set in wandb.config. Inside the training loop, metrics are logged via wandb.log({"accuracy": acc, "loss": loss}). For model weights, gradients, or input data, wandb.watch() can be used to automatically log histograms. The run ends with wandb.finish() or upon script termination.
4.3 Running and Comparing Experiments
After launching multiple runs (e.g., with different hyperparameter sets), the user views the project dashboard in the browser. Runs can be filtered by tags, config values, or run name. The user selects runs to compare on the same chart, uses parallel coordinates to explore parameter interactions, and can create reports that compare the performance across the selected runs.
4.4 Exporting and Sharing Results
Logged data can be exported as CSV, JSON, or via the API. Reports can be shared as a link (public or team‑restricted) or downloaded as PDF. For publication, users can attach a permanent report URL to a paper or presentation. Artifacts (models, datasets) can be downloaded manually or programmatically using wandb.run.use_artifact().
5.1 Official Documentation and Tutorials
Weights & Biases maintains an extensive documentation site covering all features, API references, and integration guides. Tutorials include Jupyter notebook examples, video walkthroughs, and sample projects for common frameworks. The documentation is versioned to match SDK releases and is continuously updated.
5.2 Community Forums and User Groups
The user community is supported through a dedicated forum (community.wandb.ai), a Slack workspace, and a GitHub repository for issue tracking and feature requests. Weights & Biases hosts regular webinars, meetups, and hackathons. The platform also features a Gallery of public projects showcasing best practices and novel use cases.
5.3 Pricing and Licensing Model
Weights & Biases offers a tiered pricing model. The Free tier provides unlimited runs, project sharing, and basic features for individual users or small teams. Team and Enterprise plans add advanced functionality such as augmented computing units for Sweeps, dedicated storage for Artifacts, role‑based access controls, single sign‑on, and priority support. The platform uses a software‑as‑a‑service (SaaS) model; no on‑premises deployment option was available as of 2025. Educational users and open‑source projects may qualify for free access to Team features.