Weights & Biases

Weights & Biases (W&B, or wandb after its Python package) is a developer platform for machine learning operations (MLOps) that records, visualizes, and compares model training runs. A few lines in a PyTorch or other training script send each run's configuration, such as the learning rate and batch size, along with its metrics, GPU statistics, and output files to a hosted dashboard, where teams compare experiments, reproduce results, and trace which data and settings produced a model. Beyond experiment tracking, W&B covers hyperparameter tuning, dataset and model versioning, a model registry, and, through its Weave product, tracing and evaluation of large language model (LLM) applications. The GPU cloud provider CoreWeave completed its acquisition of Weights & Biases in May 2025, and the W&B documentation now lives on the CoreWeave docs site.

Core Features#

• Experiments: wandb.init() starts a run and stores its configuration, and wandb.log() records metrics, images, and tables at each step. The SDK also samples system metrics such as GPU utilization and memory, which exposes data-loading bottlenecks.
• Sweeps: Automated hyperparameter searches using grid, random, or Bayesian methods, with agents that run trials in parallel on many machines and optional early termination of weak runs.
• Artifacts: Versioned datasets, model checkpoints, and other files, with lineage graphs that show which run consumed or produced each version.
• Registry: A curated catalog of approved model and dataset versions that teams promote toward deployment.
• Reports: Shareable documents that combine live charts from runs with written analysis.
• Weave: Tracing, evaluation, and cost tracking for LLM and agent applications.

A sweep separates the search from the training: a controller holds the search state and proposes configurations, and any number of agents pull those configurations, train, and report the target metric back.

Using W&B With Ultralytics YOLO#

Ultralytics YOLO ships a built-in W&B integration that is off by default. Once it is enabled in the Ultralytics settings, every training run logs its full argument list, per-epoch losses and learning rates, validation metrics such as precision, recall, and mean average precision (mAP), and the plots it saves, and at the end it uploads best.pt as a model artifact along with interactive precision-recall and confidence curves. Those best.pt artifacts can later be imported into Ultralytics Platform for prediction, export, or deployment.

import os

from ultralytics import YOLO, settings

# Log locally without an API key; upload later with "wandb sync", or run "wandb login" and drop this line
os.environ["WANDB_MODE"] = "offline"

# Turn on the built-in W&B callback (requires "pip install wandb"; saved to the Ultralytics settings file)
settings.update({"wandb": True})

# project and name become the W&B project and run names
model = YOLO("yolo26n.pt")
model.train(data="coco8.yaml", epochs=5, project="yolo26-experiments", name="baseline")

Training runs that share a project and differ in name appear side by side in one W&B project, so runs with different model sizes, image sizes, or augmentation settings can be compared on one dashboard.

W&B vs. Related Terms#

• Model weights and biases: The platform's name refers to the learnable parameters of a neural network. Weights scale the inputs to each neuron, biases shift its output before the activation function, and backpropagation updates both. The platform can log their distributions and gradients during training.
• MLflow: An open-source platform for tracking, packaging, and registering models that teams often self-host. W&B is primarily a managed service with self-hosted options for enterprises.
• TensorBoard: A local visualization tool for training logs from PyTorch and TensorFlow, without W&B's hosted collaboration, sweeps, or registry.