weights-and-biases

Call wandb.init with your project name and config, then call wandb.log with metrics like loss and accuracy inside your training loop. Finish with wandb.finish, and optionally save checkpoints with wandb.save or log them as artifacts.

Define a sweep config with a search method (grid, random, or bayes), a target metric, and parameter distributions, then create it with wandb.sweep. Launch one or more agents with wandb.agent pointing to your training function to execute trials.

Yes, set report_to="wandb" in TrainingArguments and the HuggingFace Trainer automatically logs metrics, evaluation results, and checkpoints to W&B. You can also add custom WandbCallback subclasses for additional logging.

Yes, set the WANDB_MODE environment variable to "offline" before initializing your run. Metrics are stored locally and can be uploaded later using the wandb sync command on the run directory.

wandb.save uploads files associated with a single run, while Artifacts create versioned, standalone objects with lineage tracking, aliases, and deduplication. Artifacts are the recommended approach for datasets and models that need versioning or registry linking.

Bayesian optimization is recommended for expensive training runs since it learns from previous trials and is most sample-efficient. Use grid search for few discrete parameters needing full coverage, and random search for quick exploration across many parameters.