> ## Documentation Index
> Fetch the complete documentation index at: https://docs.veri.studio/llms.txt
> Use this file to discover all available pages before exploring further.

# Weights & Biases

> Connect your W&B account once and every training job streams loss, reward, and throughput charts to your W&B project automatically.

Veri has a built-in [Weights & Biases](https://wandb.ai) integration for training observability. Connect your W\&B account once, and from then on every training job logs its metrics (loss, reward, learning rate, throughput) to your own W\&B project while it runs. No per-job configuration and no key in your code.

Each job's W\&B run link is surfaced everywhere the job appears:

* **Dashboard**: a "View in W\&B" button on the job page.
* **API / SDK**: the `wandb_run_url` field on the training job object.

## Connect your account

Grab an API key from [wandb.ai/authorize](https://wandb.ai/authorize), then connect it one of three ways:

<Tabs>
  <Tab title="CLI">
    ```bash theme={null}
    veri wandb set
    # W&B API key (input hidden): ...
    ```

    Optionally pick the default W\&B project runs log to:

    ```bash theme={null}
    veri wandb set --project my-experiments
    ```

    Check or remove the connection any time:

    ```bash theme={null}
    veri wandb status
    veri wandb remove
    ```
  </Tab>

  <Tab title="Dashboard">
    Open **Settings → Integrations**, paste your W\&B API key, and optionally set a default project. Save, and you're done.
  </Tab>

  <Tab title="SDK">
    ```python theme={null}
    from veri_sdk import Client

    client = Client()
    client.integrations.set_wandb(api_key="...", project="my-experiments")

    client.integrations.get_wandb()     # {"configured": true, "project": "my-experiments"}
    client.integrations.delete_wandb()  # disconnect
    ```
  </Tab>
</Tabs>

## What gets logged

Once connected, managed LLM training jobs (`grpo`, `sft_text`, `dpo`) report to W\&B automatically:

* **Project**: your configured default, or `veri-training` if you didn't set one.
* **Run name**: the job's `output_name`.
* **Run ID**: the Veri job ID, so runs and jobs correlate one-to-one.
* **Metrics**: everything the trainer emits per step (loss, reward, grad norm, learning rate, tokens/s).

The same metrics also stream to Veri's native charts on the job page. W\&B is additive, not a replacement, and jobs behave identically if W\&B is ever unreachable: training never fails because of a logging integration.

<Note>
  `sft_video_gen` jobs don't report to W\&B yet; their metrics are available on the Veri job page.
</Note>

## Custom scripts

For [custom-script jobs](/training/custom-script), Veri injects `WANDB_API_KEY`, `WANDB_PROJECT`, and `WANDB_RUN_ID` into your container's environment. A bare `wandb.init()`, or any framework with W\&B support (verl, OpenRLHF, slime, TRL), lands in your account with no code change. Your script's own `env` values take precedence if you set the same variables yourself.

After the run, Veri scans the script output for the W\&B run link and attaches it to the job as `wandb_run_url`, same as managed jobs.

## How your key is handled

* Stored encrypted at rest, scoped to your account. The API never returns it.
* Delivered to the training machine at runtime over a job-scoped authenticated channel; it is never written into job config files or logs.
* `veri wandb remove` disconnects immediately: jobs submitted afterwards no longer receive the key.