weights-and-biases

Call wandb.init with your project name and config, then use wandb.log inside your training loop to record loss, accuracy, and learning rate per step or epoch. Finish with wandb.finish and optionally upload checkpoints via wandb.save or 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 Trainer automatically logs metrics, evaluation results, and checkpoints to W&B. You can also add custom WandbCallback subclasses for additional logging.

Artifacts version any file or directory (datasets, models, predictions) with automatic lineage between runs. The Model Registry is a curated layer where model artifacts are linked and promoted through stages like staging and production using aliases.

Yes, set WANDB_MODE=offline before wandb.init to log runs locally without network access. Later, run wandb sync on the run directory to upload the logged data to the W&B servers.

Bayesian optimization is recommended for expensive runs because it learns from previous trials and samples promising parameter regions, making it the most sample-efficient method. Combine it with hyperband early termination to stop underperforming runs and save compute.