Link to this sectionLabelbox Integration#

Ultralytics Platform reads Labelbox NDJSON exports directly. There is no key to paste, no connection to keep alive, and no conversion script to maintain — export from Labelbox, upload the file, and your labels arrive as a normal Platform dataset.

Link to this sectionImport from Labelbox#

1. Export from Labelbox. Open the labeling project or catalog you want to bring over and go to the Data Rows tab. Select All, click Export data, choose the fields to include, and confirm. Exporting from Catalog rather than a project? Enable Export labels from project and pick the project, or the file arrives with no annotations at all.
2. Download the file. Labelbox runs the export as a background job. Watch for it in Notifications and click Download when it finishes to save the .ndjson file.
3. Upload to Platform. Create a new dataset and upload the .ndjson file, or paste a direct link to it. You can also start from Settings > Integrations > Labelbox.
4. Train. Once processing finishes, edit the annotations and train exactly like any other Platform dataset.

Unlike CVAT and Label Studio, there is no format to choose — Labelbox's own export is the supported one.

Link to this sectionExport with the SDK#

Labelbox also offers the export task in its Python SDK, which is the easier route for a repeatable pipeline. Stream the task and write one JSON object per line:

import json

import labelbox

client = labelbox.Client(api_key="YOUR_LABELBOX_API_KEY")
export_task = labelbox.ExportTask.get_task(client, "YOUR_EXPORT_TASK_ID")
export_task.wait_till_done()  # streaming a task that isn't COMPLETE raises

with open("dataset.ndjson", "w") as f:
    f.writelines(json.dumps(data_row.json) + "\n" for data_row in export_task.get_buffered_stream())

Link to this sectionWhat Gets Imported#

Platform recognizes the Labelbox format on its own and maps bounding boxes and polygons to the matching YOLO task type:

Each object's name becomes the class name — so a "name": "Dog" annotation imports as the class Dog, not its lowercase value — and pixel coordinates are normalized against the media_attributes dimensions in the export. A catalog export with no annotations imports as an unlabeled image dataset, ready to label in Platform's annotation editor.

A single data row looks like this, trimmed to the fields Platform reads:

{
    "data_row": {
        "external_id": "000000000307.jpg",
        "row_data": "https://storage.labelbox.com/...?Expires=...&Signature=..."
    },
    "media_attributes": { "height": 480, "width": 640, "mime_type": "image/jpeg" },
    "projects": {
        "<project-id>": {
            "labels": [
                {
                    "annotations": {
                        "objects": [
                            {
                                "name": "Dog",
                                "annotation_kind": "ImageBoundingBox",
                                "bounding_box": { "top": 200.0, "left": 407.0, "height": 172.0, "width": 176.0 }
                            }
                        ]
                    }
                }
            ]
        }
    }
}

Everything else in the export — embeddings, metadata_fields, attachments, project_details — is ignored. The embeddings are worth deselecting when you export: in a sample export they were around four-fifths of the file, against a fifth for the fields Platform actually reads.