How do I set up W&B alerts and notifications? - Weights & Biases Documentation

W&B supports two types of alerts: automated alerts configured in User Settings (run finished, run crashed) and programmatic alerts fired from inside your training code with wandb.alert(). Automated alerts via User Settings Go to wandb.ai/settings and scroll to the Alerts section. You can enable:

Run finished — notifies you when any run in your projects completes successfully.
Run crashed — notifies you when a run exits without calling wandb.finish().

Notifications can be delivered to email, Slack, or both. To connect Slack, click Connect Slack in the Alerts settings and follow the OAuth flow. Note: Run Finished alerts do not fire in Jupyter notebook environments—use wandb.alert() in notebooks instead. Programmatic alerts with wandb.alert() Send an alert from anywhere in your training loop:

import wandb

wandb.init(project="my-project")

for step in range(1000):
    loss = train_step()
    wandb.log({"loss": loss}, step=step)

    if loss > 10.0:
        wandb.alert(
            title="Loss spike detected",
            text=f"Loss jumped to {loss:.2f} at step {step}",
            level=wandb.AlertLevel.WARN,
        )

Alert levels are wandb.AlertLevel.INFO, wandb.AlertLevel.WARN, and wandb.AlertLevel.ERROR. The level controls the visual appearance of the notification but does not affect delivery. Preventing alert spam with wait_duration If your training loop could trigger the same alert repeatedly, use wait_duration to throttle:

import datetime

wandb.alert(
    title="High loss",
    text=f"Loss is {loss:.2f}",
    level=wandb.AlertLevel.WARN,
    wait_duration=datetime.timedelta(minutes=30),  # at most one alert per 30 min
)

With wait_duration set, subsequent calls with the same title are silently skipped until the duration has elapsed. Team-level alerts Automated alerts in User Settings apply to your personal account. For team-wide alerting (for example, notifying a shared Slack channel when any team member’s run crashes), each team member must connect their own Slack in their User Settings, or use a shared service account with Slack connected. Team-level alert configuration is not available directly in the UI today—wandb.alert() in code with a shared Slack webhook is the recommended workaround.

Alerts Experiments Runs

Documentation Index