Weights & Biases

• You are training ML models and want to track metrics, configs, and artifacts automatically across runs.
• You need to compare experiments across hyperparameters, architectures, or datasets in a shareable dashboard.
• You want to run automated hyperparameter sweeps (grid, random, or Bayesian) and let the assistant manage the search.
• You are working on a team and need a central model registry with lineage and versioning.

• You are not doing ML training and have no need for experiment tracking or model management.
• Your environment has no internet access and you cannot self-host a W&B server.
• You only need lightweight local logging — a simple CSV or TensorBoard may be sufficient.

Initialize runs with wandb.init, log metrics and hyperparameters with wandb.log, and have every run automatically streamed to your W&B project dashboard in real time.

Define sweep configurations using grid, random, or Bayesian strategies. The assistant can set up the sweep, launch agents, and help you interpret results to find the optimal configuration.

Log datasets, model checkpoints, and other files as typed artifacts. W&B tracks full lineage — which dataset produced which model — so you can reproduce any result.

Tag artifact versions with aliases like best or production and link them to a centralized model registry. Supports team-level governance and deployment promotion workflows.

Works natively with PyTorch, TensorFlow/Keras, HuggingFace Transformers, and PyTorch Lightning via callbacks and loggers — minimal code changes required.

Every run gets a shareable URL. Build reports combining charts, runs, and markdown for team review, and organize experiments with tags, groups, and project-level visibility controls.

Instrument a new training script to log loss, accuracy, and hyperparameters to W&B, then use the dashboard to compare it against future runs and identify regressions.

Define a Bayesian sweep over learning rate, batch size, and dropout. Launch multiple agents to run trials in parallel and surface the best configuration automatically.

Save model checkpoints as versioned artifacts after each training run, alias the best-performing checkpoint, and link it to the model registry for downstream deployment.

Add report_to="wandb" to HuggingFace TrainingArguments and have the assistant scaffold the full setup — project init, run naming, and post-training artifact logging.