lacing.adapters.label_studio#

Label Studio JSON adapter.

Label Studio is a popular labeling tool whose canonical export shape is a JSON array of tasks, each with one or more annotations containing typed results. We focus on time-interval results (the Audio, Video, and TimeSeries labeling controls) since lacing is an interval-annotation system.

Mapping#

Label Studio ↔ lacing data.audio / data.video / data.url ↔ MediaRef.asset_id result.from_name ↔ Tier.name result.value.{start, end} (seconds) ↔ MediaRef.interval result.value.labels (list of strings) ↔ body['labels'] annotation.completed_by (or user) ↔ Provenance.was_attributed_to result.origin (“prediction”/…) → confidence heuristic result.id → preserved in body result.type → preserved in body

Lossy on dump#

  • Per-result Label Studio metadata (image_rotation, original_width, etc.) is dropped — they’re irrelevant for time intervals.

  • Predictions vs annotations distinction is collapsed: lacing tracks AI vs human via confidence and provenance.was_generated_by, and we set Label Studio’s origin to manual on dump.

  • Multi-annotator support: only one annotation per task is emitted on dump (the first), since lacing’s per-annotation provenance fits Label Studio’s per-task annotations list awkwardly.

Spec: https://labelstud.io/guide/export.html

Functions

dump(store[, target, pretty])

Serialize store as a Label Studio JSON export.

load(source, *[, rate, asset_id, attribution])

Load a Label Studio JSON export into a MemoryStore.
lacing.adapters.label_studio.dump(store, target=None, *, pretty=True, **_kwargs)[source]#

Serialize store as a Label Studio JSON export.

Annotations are grouped by MediaRef.asset_id — each asset becomes one task with a single annotations[0] entry collecting every region. confidence < 0.5 flips the result into Label Studio’s predictions slot rather than annotations.

Return type:

bytes | None

lacing.adapters.label_studio.load(source, *, rate=24000, asset_id=None, attribution=None, **_kwargs)[source]#

Load a Label Studio JSON export into a MemoryStore.

Parameters:

  • source (str | bytes | PathLike) – Path, bytes, or JSON string. Accepts both an array of tasks and a single task dict.

  • rate (int) – Quantization rate.

  • asset_id (str | None) – Override the resolved asset for every annotation.

  • attribution (str | None) – Override per-annotation attribution.

Return type:

IntervalAnnotationStore