weights-and-biases_skill

Overview

This skill integrates Weights & Biases (W&B) into ML workflows to track experiments, visualize training in real time, run hyperparameter sweeps, and manage model artifacts and registry. It provides ready-to-use patterns for logging metrics, saving checkpoints, tracking datasets, and collaborating across teams. Use it to add MLOps-grade observability and reproducibility to research and production training pipelines.

How this skill works

The skill instruments training scripts to automatically log configs, metrics, media, and system telemetry to the W&B service or local offline storage. It supports simple API calls (wandb.init, wandb.log, wandb.Artifact) that attach run metadata, upload artifacts, and stream dashboards. For large-scale search, it uses sweep configurations and wandb.agent to orchestrate grid, random, or Bayesian hyperparameter optimization.

When to use it

• When you need reproducible experiment tracking and searchable run history.
• When you want live dashboards to monitor training metrics, gradients, and media.
• When optimizing hyperparameters across many trials with automated sweeps.
• When you must version datasets, models, and other artifacts with lineage.
• When collaborating in teams and sharing run reports or production models.

Best practices

• Initialize runs with descriptive project, name, tags, and group to organize experiments.
• Log hyperparameters and config at run start so runs are fully reproducible.
• Upload important checkpoints and final models as Artifacts and use aliases like 'best' or 'production'.
• Log system metrics (GPU, CPU, memory) and code version (git_commit) for troubleshooting.
• Use offline mode for unstable networks and sync runs later to avoid data loss.

Example use cases

• Track training and validation curves for a ResNet50 image classifier, with images and confusion matrices attached for review.
• Run a Bayesian sweep to maximize validation accuracy across learning rate, batch size, and optimizer choices.
• Log a Transformer fine-tuning run from Hugging Face Trainer with automatic W&B reporting and model checkpointing.
• Publish a model to the W&B model registry with metadata and link it to a production deployment.
• Share interactive reports combining runs, charts, and narrative text for team reviews or reproducible publications.