Tool Reference

Processing

Prepare training-ready data with tools for merging datasets, removing unlabeled items, applying augmentation transforms, and exporting to 20+ industry formats. All processing runs asynchronously on serverless GPU infrastructure — use job tracking tools to monitor progress.

merge_datasets

Async Job

Merge multiple datasets of the same type into one. Automatically detects dataset type (classification, detection, segmentation, keypoints, OCR, video, 3D, vision-language, image-to-image) and routes to the appropriate merge function with intelligent class reindexing.

Parameters

dataset_namestring

Name for the merged output dataset.

Required

additional_dataset_pathslist[string]

Dataset names to merge into the target.

Required

output_folderstring

Custom output folder in storage.

Optional

class_mappingdict

Per-dataset class name remapping, e.g. {"dataset-a": {"car": "vehicle"}}.

Optional

branch_namestring

Branch to operate on. Defaults to "main".

Optional

Returns

Job ID. When complete, the merged dataset is available under the specified name with all classes reindexed.

Example prompt

“Merge "cars-v1" and "cars-v2" into a new dataset called "cars-combined"”

clean_dataset

Async Job

Remove items (images, videos, or 3D assets) that have no labels from a dataset. Automatically detects dataset type and routes to the appropriate handler. Useful for removing unlabeled data before training.

Parameters

dataset_namestring

Name of the dataset to clean.

Required

branch_namestring

Branch to operate on. Defaults to "main".

Optional

Returns

Job ID. When complete, a new version of the dataset is created with unlabeled items removed.

Example prompt

“Clean my retail-detection dataset by removing all unlabeled images”

augment_dataset

Async Job

Augment any dataset type by applying transformations. Can target specific classes and control the number of augmented copies per image.

Parameters

dataset_namestring

Name of the dataset to augment.

Required

augmentation_typeslist[string]

Transforms to apply, e.g. ["horizontal_flip", "rotate", "blur"].

Required

augmentation_paramsdict

Per-transform parameters, e.g. {"rotate": {"angle": 15}}.

Optional

num_augmentationsint

Number of augmented copies per image. Defaults to 1.

Optional

classeslist[string]

Only augment images containing these classes.

Optional

branch_namestring

Branch to operate on. Defaults to "main".

Optional

Returns

Job ID. When complete, augmented images and labels are added to the dataset.

Example prompt

“Augment my PPE dataset with horizontal flip, rotation, and brightness adjustments — 3 copies per image”

export_dataset

Async Job

Export a dataset to a specific format for download. Checks the export library for cached exports first. Supports 20+ formats across all dataset types.

Parameters

dataset_namestring

Name of the dataset to export.

Required

export_formatstring

Target format (e.g. "yolo", "coco", "voc").

Required

splitslist[string]

Splits to include. Defaults to ["train", "val", "test"].

Optional

branch_namestring

Branch to export. Defaults to "main".

Optional

force_newboolean

Skip cache and create a fresh export.

Optional

Returns

Job ID. When complete, returns a download URL for the exported archive.

Example prompt

“Export my safety-helmet dataset in YOLO format”

Supported export formats

Detection

yolococovockitti

Segmentation

yolo_segcoco_segmasks

Keypoints

coco_keypointsyolo_pose

Classification

foldercsvclassification

Vision-Language

jsoncaptionvqa

Image-to-Image

pairedpix2pix

Video

tracking

plyobjstlgltf

Hub

huggingface

Augmentation transforms

horizontal_flipvertical_fliprotatebrightnesscontrastsaturationblurnoisecropresize