Fine-tuning

Before you launch

Create a project, ingest ROS 2 MCAP, LeRobot 2.1, or LeRobot v3 source data, curate the episode selection, and pin a DatasetVersion. A version freezes the exact episode revisions, source artifact roots, schema/mapping/calibration revisions, normalization statistics, and source commits used by training.

Training never follows a mutable repository branch or a Hugging Face dataset name. Editing episodes or changing the dataset definition creates new history; an existing DatasetVersion continues to resolve the same byte and semantic closure.

Configure the job

Select one of the supported backends:

• ACT
• SmolVLA
• Pi0.5
• SARM reward training

SmolVLA and Pi0.5 accept an external foundation-model origin. ACT and SARM use their fixed platform model and reject a custom model_id.

Choose an immutable dataset version, then map the model’s camera slots to the adapter-pinned camera features shown by the launcher. Those descriptors include the annotation key, source stream, timebase, and frame-to-timestamp mapping; the job is rejected if a mapping does not belong to the selected version.

Finally choose the GPU, region, duration, batch size, and optional validated model hyperparameters.

What happens at execution

Before a pipeline run is created, Qualia resolves the DatasetVersion into an immutable native read-plan identity. Each reader attempt receives a separate short-lived credential bound to:

• the active op-run and retry attempt;
• the owning account;
• the exact DatasetVersion UUID;
• the plan version, layout, manifest root, and plan digest.

The worker verifies the canonical plan header, every digest-bound page, each stored object envelope, chunk identity, slice range, reconstructed artifact, and final file digest before training consumes it. A cancelled, failed, expired, non-reader, or superseded attempt cannot mint fresh object URLs.

Monitor and cancel

The job view is backed by durable pipeline operations rather than a synthetic preprocessing state machine. It shows credit reservation, optional SARM progress computation, training, model publication, and credit finalization, with op logs and telemetry attached to the active attempt.

Cancellation revokes the attempt’s dataset read credential, stops remaining pipeline work, and finalizes credits. Already-issued object URLs are bounded, whole-object GET capabilities with a short expiry.

Reproduce a run

Record the job ID, DatasetVersion UUID, read-plan digest, foundation-model revision, hyperparameter document, and resulting model artifact/commit. The dataset edit/audit trail remains in Git, while the DatasetVersion and model artifact provide immutable data and output anchors.

For programmatic use, see the Finetune SDK resource and Data SDK resource.