Logging metrics

The observability dashboard captures the most important plots and metadata you’d care about during training, and its Metrics tab charts every scalar the underlying framework logs through wandb.log — with any metric provider. Use Weights & Biases or Trackio when you also want those numbers in an external tracker.

Dashboard only

This is the default: SlimeRecipe and MilesRecipe start with metrics=DashboardMetricConfig() (a few model recipes override it, e.g. Qwen3_6_27B_Recipe_Agentic ships with Trackio), which sends the framework’s metrics to the dashboard and nowhere else — no account, API key, or extra server. Set it explicitly to name the project or group:

from modal_dojo import DashboardMetricConfig, Qwen3_5_4B, Qwen3_5_4B_Recipe, TrainConfig

config = TrainConfig(
    model=Qwen3_5_4B(),
    dataset=my_dataset,
    recipe=Qwen3_5_4B_Recipe(
        # ...
        metrics=DashboardMetricConfig(project="my-rl-project"),  # optional; this is the default provider
    ),
)

run = config.launch()

Open the run in the dashboard and switch to the Metrics tab: keys are grouped by prefix like W&B panels (train/, rollout/, perf/, …), the search box filters every group, and charts refresh while the run trains. Only finite scalars are kept (nested dicts flatten to a/b; images, tables, and strings are dropped).

Pass metrics=None to turn metric logging off entirely — the framework runs without --use-wandb and nothing is stored.

Weights & Biases

When you need everything W&B offers (media, tables, cross-project reports), pass a WandbConfig. The framework logs to W&B as usual and the same scalars also appear in the dashboard’s Metrics tab.

First, you’ll need to create a Modal Secret with your API key:

modal secret create wandb-secret WANDB_API_KEY=<your-api-key>

Then, just pass it in your training recipe:

from modal_dojo import Qwen3_5_4B, Qwen3_5_4B_Recipe, TrainConfig, WandbConfig

config = TrainConfig(
    model=Qwen3_5_4B(),
    dataset=my_dataset,
    recipe=Qwen3_5_4B_Recipe(
        # ...
        metrics=WandbConfig(
            project="my-rl-project",
            group="lr-sweep",  # optional: organize related runs
        ),
    ),
)

run = config.launch()

See the reference page for the full list of parameters.

When launching a hyperparameter sweep, the group parameter is especially useful to overlay multiple runs’ reward curves.

Trackio

Trackio is an open-source W&B alternative which you can deploy on Modal or on a Hugging Face Space. You can even self-host! Modal Dojo installs it in the training image and routes the framework’s metric calls to it whenever a recipe uses TrackioConfig; scalars also appear in the dashboard’s Metrics tab.

Deploy on Modal

You can host a Trackio server on Modal:

from modal_dojo import TrackioConfig

metrics = TrackioConfig.deploy_to_modal(project="my-rl-project")

Just like the main dashboard, the Trackio dashboard is unauthenticated unless you set a password:

modal-dojo set-password

Note that unlike the main dashboard, this will not redeploy the Trackio dashboard. I.e., you’ll have to rerun deploy_to_modal() after changing it.

Deploy on a Hugging Face Space

Simply specify a space_id and optionally a bucket_id:

from modal_dojo import Qwen3_5_4B, Qwen3_5_4B_Recipe, TrainConfig, TrackioConfig

config = TrainConfig(
    model=Qwen3_5_4B(),
    dataset=my_dataset,
    recipe=Qwen3_5_4B_Recipe(
        # ...
        metrics=TrackioConfig(
            project="my-rl-project",
            space_id="my-org/training-metrics",
            bucket_id="my-org/training-metrics",  # optional
        ),
    ),
)

Self-hosted

You’ll need to create a Modal Secret with your API key:

modal secret create trackio-write-token TRACKIO_WRITE_TOKEN=<your-api-key>

Then, it’s as easy as:

metrics = TrackioConfig(
    project="my-rl-project",
    server_url="https://trackio.example.com",
    modal_secret_name="trackio-write-token",
)

See the reference page for all parameters.