Experiments
The standard for experiment tracking — automatically recording and comparing the metrics, configs, and results of every training run.
With a few lines of code, automatically log everything about an experiment — hyperparameters, loss, accuracy, GPU utilization — and compare runs side by side on interactive dashboards. Leave a fully reproducible record of who trained what, when, and with which settings.
What Is Experiments
As you develop AI models, dozens to hundreds of training runs pile up, and it's easy to lose track of which configuration worked best. With a single `wandb.log()` line, W&B Experiments automatically records metrics, hyperparameters, system metrics, and artifacts — turning your experiments into a system of record you can compare and reproduce at any time.
Key Capabilities
Automatic logging
Records metrics, hyperparameters, GPU/CPU stats, and even gradients automatically
Interactive dashboards
Compare, filter, and visualize runs side by side (loss curves, confusion matrices, and more)
Reproducibility
Captures code version, environment, and dataset together for full reproduction
Framework integrations
Built-in support for PyTorch, TensorFlow, Keras, Hugging Face, XGBoost, and more
What Gets Logged
| Item | Contents |
|---|---|
| Metrics | Time series of loss, accuracy, and custom metrics |
| Configuration | Hyperparameters · config |
| System | GPU/CPU utilization, memory, and power |
| Artifacts | Checkpoint, dataset, and output versions |
What It's Used For
Tracking model development
Keep an experiment history to find the best configuration fast
Team collaboration
Share and review experiment results with the team
Reproduction & audit
Secure the reproducibility required in regulated and research settings