Weights & Biases (W&B)

W&B is the "Github for ML models". It tracks every run, hyperparameter, and artifact. 2025 brings W&B Inference and Weave.

When to Use

• Visualization: Live loss curves during training.
• Collaboration: Sharing reports ("The loss spiked here") with teammates.
• Model Registry: Managing lifecycle.

Core Concepts

W&B Run

A single experiment.

Sweeps

Hyperparameter optimization engine.

Weave

A toolkit for logging and debugging interactive GenAI applications (like Tracing).

Best Practices (2025)

Do:

• Use Weave: For LLM apps. It replaces simple text logging.
• Log Model Artifacts: Save best_model.pt to W&B, not local disk.

Don't:

• Don't leak secrets: W&B logs configs. Ensure API keys aren't in your config object.

References

• Weights & Biases