Weights & Biases

Machine learning experiment tracking and model development platform with run comparisons, hyperparameter optimization, evaluation tables, and collaborative workflows.

About Weights & Biases

Weights & Biases helps developers record and compare machine learning experiments. Its Models offering captures training information and presents results in shared workspaces, helping teams inspect performance changes and understand which configurations produced them.

The platform includes experiment tracking, hyperparameter search, evaluation tables, and tools for collaboration. It integrates with widely used training frameworks so developers can add logging to their existing model-development code.

Weights & Biases Models is presented within CoreWeave Forge, and cloud pricing is published through that platform. Private-hosting options have separate terms. This profile focuses on the model-development and experiment-management capabilities relevant to machine learning teams.

Key Features

Experiment Tracking: Record training metrics, hyperparameters, and run information.
Run Comparison: Compare model experiments and investigate performance changes.
Hyperparameter Sweeps: Automate searches using Bayesian optimization, grid search, or random search.
Evaluation Tables: Inspect aggregate scores alongside example-level inputs and outputs.
Shared Visualizations: Explore results and communicate findings through collaborative workspaces.
Deployment Choices: Access shared cloud, dedicated cloud, and customer-managed options according to the offering.

Pricing

Cloud Free: $0/month. Up to 5 model seats, 5GB storage/month, and 1GB Weave ingestion/month.
Cloud Pro: Starts at $60/month, billed monthly upfront. Up to 10 model seats, 100GB storage/month, and 1.5GB Weave ingestion/month. Includes a 30-day trial; intended for organizations with fewer than 50 employees.
Cloud Usage: Pro storage overages $0.03/GB; Weave ingestion $0.10/MB. Usage is billed monthly in arrears. Inference and sandboxes cost extra; a temporary $5/month inference credit is advertised. Added seats are prorated; removed seats receive no credits. Additional seat pricing is unclear.
Cloud Enterprise: Custom pricing; invoiced annually upfront. Published additional Weave ingestion rate: $0.20/MB.
Academic License: Free Pro license for eligible academic research unrelated to a for-profit entity. Includes 200GB storage, up to 25GB monthly Weave ingestion, and up to 100 seats. Extra storage: $0.03/GB/month.
Private Personal: $0/month. One user; personal projects only. Corporate use is prohibited.
Private Advanced Enterprise: Contact sales. Custom plan with an enterprise trial available.

Use Cases

Compare model-training runs and their hyperparameters.
Automate searches for promising training configurations.
Investigate regressions through example-level evaluation results.
Coordinate model-development work across a shared team workspace.

Pros & Cons

Pros:

Combines run tracking, visual comparison, and parameter search.
Integrates with several established training frameworks.
Supports multiple deployment approaches.

Cons:

Cloud pricing and private-hosting terms must be evaluated separately.
The free private-hosted Personal plan does not permit corporate use.

Integrations

PyTorch, TensorFlow, Keras, scikit-learn, XGBoost, Hugging Face Transformers, Lightning

FAQ

What is Weights & Biases used for in machine learning?

Weights & Biases helps developers track training experiments, compare runs, inspect model performance, and collaborate on model development.

What are Weights & Biases sweeps?

Sweeps automate hyperparameter searches. The Models offering documents Bayesian optimization, grid search, and random search approaches.

Which training information can Weights & Biases record?

It can record metrics, hyperparameters, and other experiment information supplied through its SDK and framework integrations.

What are evaluation tables used for?

Evaluation tables help compare overall scores with example-level model inputs, outputs, and scores, making it easier to investigate changes between model versions or checkpoints.

Does Weights & Biases work with existing training code?

Yes. Its official materials show integrations with frameworks including PyTorch, TensorFlow, Keras, scikit-learn, XGBoost, Hugging Face Transformers, and Lightning.

Can Weights & Biases run in a customer-managed environment?

Yes. Its deployment documentation describes customer-managed cloud and on-premises options alongside hosted services. Licensing and available controls depend on the offering.

How is Weights & Biases Models related to CoreWeave Forge?

The official Models page directs visitors to the Weights & Biases Models offering within CoreWeave Forge, while the original private-hosting page links to Forge for cloud pricing.

Categories:

Dev Tools Machine Learning Tools