Projects & datasets

A project groups everything Base holds for one solarian project. Its datasets are the recordings you upload, which never change, and the computed datasets you derive from them with ops, version by version.

Projects

From the web app: Projects → New Project asks for a name and an optional description. A project's page has, from top to bottom:

• Edit, for the name and description;
• Deployments — see Deployments & Secrets;
• Training and Start training — see Training & Runs;
• Datasets and New Dataset;
• Environment variables.

Deleting a project, from the Projects list, deletes its datasets too.

From the CLI: solarian base link --create creates a project named after the current directory and links it. There are no other project commands.

Upload a dataset

From the web app: on the project page, New Dataset → Upload. Pick the archive, give it a name, and press Upload Dataset.

The format comes from the archive's contents, not its name; .7z works too. An upload can be up to 50 GB, within your plan's storage (5 GB on Free). See Dataset Formats for what each format holds.

Datasets lists every dataset of every project.

View a dataset

A dataset's page plays one episode at a time:

• Observations — every camera, in sync, with tracked boxes and masks drawn on top when the dataset has them, and a 3D gripper view when it has poses;
• Actions — one chart per action and state series, following the playhead;
• the sidebar's Data chooses which series and cameras are shown, and Info gives the episode, frame and FPS counts.

Delete removes the current episode from an uploaded dataset. On a computed dataset, use the delete-episodes op instead, which keeps the original.

Start training on a ready, uploaded dataset opens the training dialog with it already selected.

Computed datasets

A computed dataset is a tree of versions over one uploaded Solarian dataset. Each version applies one op to its parent, so the source stays untouched and any step can be redone.

From the web app: New Dataset → Computed from… on the project page, or New computed dataset on a Solarian dataset's page. Then, in Versions:

1. Select the source or a version and press Apply op….
2. Choose the op and its parameters. Ops that work per episode offer Batch or Episode by episode.
3. Press Apply. Log shows the op's output while it runs, and Cancel stops it.

Versions are numbered in the order they were made and named after their op, such as 01-track or 02-split-pick-place-pick.

track comes from the solarian-sam2 plugin and track_point from solarian-tapnext; both write into the episode's targets/<camera>.npz without dropping each other's arrays, so one version can carry a pick box and a place point. track_point uses the running TAPNext endpoint when there is one. A project can add its own ops in <package>/dataset_ops/, scaffolded with solarian add-op; Base lists them with (project) next to their name.

Review sessions

Episode by episode opens the version for review. Base keeps a worker on it while you go through the episodes:

• for a filter such as delete-episodes: Keep ⏎ or Drop D;
• for a map such as track: watch the result, then Accept ⏎, Redo R (after clicking a new prompt on the frame), or Stop Esc.

Once every episode is decided, Close makes the version final, so ops can build on it and it can be trained on or exported.

Use a version

• Export… turns a closed version into a new, standalone uploaded dataset.

• Download gets the version as one .solarian.zip, which is what training reads.

• Drop deletes the version.

• To train on it without exporting, reference it as base://datasets/<id>@<version>:

  solarian base train umi_pick "++task.dataset_mix.0.path=base://datasets/3f2a9c1e@02-split-pick-place-pick"
  

  CopyCopied!

See also

• Dataset Formats — LeRobot, UMI and Solarian layouts.
• Training & Runs — training on Base datasets.
• Recording — producing Solarian datasets.