Export results and datasets

Keep artifacts, groups, splits, rights and provenance.

Export a completed result

The quickstart's --export writes the common JSON summary, JUnit, result envelope and exact JSON views to a new local directory. It does not overwrite previous evidence. JSON views are a versioned transport projection; their serialized bytes are not the original artifact hash input.

Call get_result to obtain original artifact names, SHA-256 values and authorized download references. Download JSONL, ZIP and other original bytes from the artifact route with the current bearer token, then compare their SHA-256 before use. run_goal(include_artifacts=True) embeds JSON views; other files appear in export_skipped with their original download path.

Dataset export is execution, not preparation

export_result and verify_result return plans. A prepared plan explicitly says execution has not started, its run is null and no export artifact has been created. Use execute_plan with the plan/digest and an idempotency key, or the typed run_goal examples on Supported tasks. Confirm export only after completed execution and existing checked files.

export_dataset produces episodes.jsonl and dataset-manifest.json for the selected domain. Select complete linked groups; a successful/failing pair cannot be split by choosing only one label. Preserve original IDs, payloads, rights, source/profile versions and train/validation/test assignments.

Versioned dataset generations

build_dataset under dataset.navigation.jsonl.v1 or dataset.can.lerobot.v3 creates an immutable DatasetManifest with exact selected episode IDs, purpose and source lineage. Inspect it using get_dataset_generation.

LeRobot output uses original named observations, not the full simulator state as a substitute. The pinned reader runtime is listed in the profile contract; it is not installed by the base package. The accepted reader round trip is retained M5 evidence, not a claim that every export environment is installed.

Training generations accept only original train records. Normalization uses train data only. Evaluation purpose does not authorize training, redistribution or a change in source rights. The service rejects incomplete groups, duplicate IDs and held-out misuse.

Reuse checked cases

find_verified_results finds compatible results for an explicit dataset and purpose. create_suite includes deliberately selected result IDs without duplicating them; get_suite reads the immutable selection. Search alone does not place foreign or held-out data into a training base.

Preparation contract ? Result/artifact contract ? Executable examples