labelpull

Pull the latest Labelbox annotations into a tidy, ontology-agnostic table.

The Labelbox SDK already exports a project's labels and streams them. What it doesn't give you is a tabular view of that deeply nested JSON, the correctness logic to pick the right label when a row was reviewed, or a workflow status that is always populated. labelpull is exactly that thin layer on top of the SDK.

Quick start

pip install 'labelpull[live]'
export LABELBOX_API_KEY=...                 # Labelbox → Workspace settings → API keys
labelpull pull <PROJECT_ID> -o labels.csv   # <PROJECT_ID> is in your project's URL

You get one row per annotation (any ontology):

global_key,data_row_id,feature_kind,feature_name,value,workflow_status,labeled_by,created_at,parent_feature_id
photo_001.jpg,clz…,label,,,Done,bot@lab.org,2026-06-05T08:00:00Z,
photo_001.jpg,clz…,radio,Species,Ficus insipida,Done,bot@lab.org,2026-06-05T08:00:00Z,
photo_001.jpg,clz…,checklist,Organs,leaf;flower,Done,bot@lab.org,2026-06-05T08:00:00Z,

The first row of each photo (feature_kind=label, empty feature_name) is a marker that the photo was reached and labelled — it carries who/when even when the annotator left it blank, so empty-but-labelled photos still show up. Ignore it if you only want answers: filter to feature_kind != "label".

Install

pip install labelpull          # offline parsing + CLI (no SDK)
pip install 'labelpull[live]'  # + the Labelbox SDK, for live pulls from the API

CLI

labelpull pull <PROJECT_ID> -o labels.csv               # everything, generic long CSV
labelpull pull <PROJECT_ID> --status Done               # only verified rows
labelpull pull <PROJECT_ID> --since 2026-06-01          # only labels created since a date
labelpull pull <PROJECT_ID> --from-export export.ndjson # offline: a UI "Export" file, no API key

--status takes ToLabel | InReview | InRework | Done. Every run prints a summary (rows, labelled count, feature kinds, latest label timestamp).

If your project is a single-classification task and you want one row per item instead of the long format, filter the CSV to your feature (e.g. keep feature_name == "Species"), or write a 10-line Adapter (see below).

Library

import labelpull

rows = list(labelpull.export("proj_id", status="Done"))   # live; needs labelpull[live]
# or, offline from a UI export:
# rows = labelpull.read_export_file("export.ndjson")

features = [f for r in rows for f in labelpull.flatten(r, "proj_id")]
labelpull.write_csv("labels.csv", labelpull.GenericAdapter(), features)
print(labelpull.summarize(rows, features))

flatten() handles radio / checklist / text classifications and bbox / polygon / line / point / mask objects (with nested classifications linked to their parent), and always selects the most recently created label so a QC-reviewed row reports the reviewer's answer, not the annotator's.

Custom output shape

GenericAdapter (the default) writes one row per feature. To collapse features into a project-specific wide table, write an Adapter — given the flattened FeatureRows, yield your own columns. SpeciesAdapter is a worked example (it pivots a Taxon radio + Organs checklist into one row per photo):

labelpull pull <PROJECT_ID> --schema species -o taxa.csv