Tool Reference

Label Operations

Refine your annotation taxonomy without manual relabeling. These tools let you rename classes for consistency, remove unwanted labels, and query label distributions — all with automatic dataset-type detection for classification, detection, segmentation, keypoints, and OCR.

modify_labels

Async Job

Modify labels in a dataset — remove unwanted classes or rename/remap class names. Automatically handles reindexing of remaining classes across all annotation files. Supports classification, detection, segmentation, keypoints, and OCR datasets.

Parameters

dataset_namestring

Name of the dataset.

Required

actionstring

"remove" to delete classes, "replace" to rename classes.

Required

labels_to_removelist[string]

Classes to remove (when action is "remove").

Optional

label_mappingdict

Old→new name mapping, e.g. {"person": "human"} (when action is "replace").

Optional

old_labelstring

Single label to rename (alternative to label_mapping).

Optional

new_labelstring

New name for old_label.

Optional

branch_namestring

Branch to operate on. Defaults to "main".

Optional

Returns

Job ID for tracking progress via get_processing_status. Creates a new dataset version — original data is preserved.

Example prompt

“Remove the classes "Scissors" and "Knife" from my airport-xray dataset”

query_dataset

Instant

Query dataset labels and images. Returns instant results from the database without running compute jobs. Use "count_labels" to get label distribution or "find_images" to locate images containing a specific class.

Parameters

dataset_namestring

Name of the dataset.

Required

actionstring

"count_labels" (default) or "find_images".

Optional

labelstring

Label to search for (required for find_images).

Optional

exclusiveboolean

If true, only return images that contain exclusively this label.

Optional

branch_namestring

Branch to query. Defaults to "main".

Optional

Returns

For count_labels: object mapping class names to counts. For find_images: list of image filenames matching the query.

Example prompt

“How many images of each class are in my retail-detection dataset?”

Behavior by dataset type

ClassificationRemoves/renames classes. Images belonging exclusively to removed classes are deleted.

DetectionRemoves/renames bounding box classes. Images are kept unless all annotations are removed.

SegmentationUpdates polygon/mask annotations. Preserves segmentation data for remaining classes.

KeypointsUpdates keypoint annotations. Preserves visibility flags and coordinates.

OCRRemoves text regions matching exact text. For replacement, updates the text content.

Non-destructive: Label operations create a new version of your dataset. Your original data is always preserved and you can revert at any time.