Label Studio

Provides 40+ pre-built annotation templates (classification, NER, bounding box, polygon, keypoint, relation extraction, etc.) that can be composed via XML-based label configuration. The frontend uses React with canvas-based rendering for spatial annotations and dynamically loads template schemas that map to backend task models, enabling users to define custom labeling interfaces without code.

Implements a pluggable next-task algorithm (in label_studio/projects/functions/next_task.py) that ranks unlabeled tasks based on sampling strategies (random, sequential, uncertainty sampling from ML predictions, consensus-based disagreement). The Data Manager API filters and sorts tasks using database queries with optional ML model predictions, enabling prioritization of high-value samples for labeling efficiency.

Implements FSM-based state transitions for tasks (label_studio/tasks/models.py or similar) where tasks move through defined states (unlabeled → in-progress → completed or skipped). State transitions are validated to prevent invalid state changes (e.g., cannot go from completed back to unlabeled). FSM is configurable per project, allowing custom state workflows.

Integrates a background job queue (likely Celery with Redis or similar) for asynchronous processing of long-running tasks (bulk import, export, ML prediction requests, annotation processing). Jobs are queued, executed by worker processes, and results are stored in the database or cache. Job status can be tracked via API.

Uses Django migrations (label_studio/migrations/) to version database schema changes and manage schema evolution. Migrations are applied sequentially during deployment, enabling rollback if needed. Supports both forward and backward migrations for schema compatibility.

Exposes comprehensive REST APIs (label_studio/projects/api.py, label_studio/tasks/api.py, label_studio/organizations/api.py, etc.) for all platform features (project management, task CRUD, annotation CRUD, user management, storage configuration, ML integration, import/export). APIs use Django REST Framework with token-based authentication and support filtering, pagination, and sorting. API documentation is auto-generated from code.

Provides an ML API (label_studio/ml/api.py) that accepts predictions from external models via REST endpoints, stores predictions in the database, and displays them as pre-filled annotations in the labeling interface. Supports both synchronous prediction requests (send task data to model, receive predictions) and asynchronous batch prediction uploads. Predictions are versioned and can be compared against ground-truth annotations for model evaluation.

Implements pluggable storage backends (label_studio/io_storages/) that connect to cloud providers via their native SDKs (boto3 for S3, google-cloud-storage for GCS, azure-storage-blob for Azure). Tasks can be imported directly from cloud buckets, and annotations can be exported back to cloud storage. Storage configuration is managed per-project with credentials stored encrypted in the database, enabling multi-cloud deployments without code changes.

Implements Django-based user and organization management (label_studio/organizations/, label_studio/users/) with role-based access control (RBAC) at project and organization levels. Tracks annotation history per task, enabling review of who labeled what and when. Supports team workspaces with per-project role assignments (annotator, reviewer, manager) and audit logging for compliance.

Implements a data import pipeline (label_studio/data_manager/api.py, label_studio/io_storages/) that accepts multiple file formats (JSON, CSV, XML, images, videos, audio, time series) and automatically detects format based on file extension or MIME type. Imported data is parsed and converted into Task objects in the database, with support for bulk import via ZIP files or cloud storage. Import progress is tracked asynchronously via background jobs.

Implements an export pipeline (label_studio/tasks/api.py, label_studio/io_storages/) that converts annotations from internal JSON format to multiple output formats (JSON, XML, CSV, COCO, Pascal VOC, YOLO, Hugging Face datasets). Export can be filtered by annotation status (completed, in-progress, skipped), annotator, or date range. Exports are generated asynchronously and can be downloaded or pushed to cloud storage.

Provides a project settings API (label_studio/projects/api.py, label_studio/projects/models.py) that allows users to configure project metadata (name, description, label configuration), sampling strategy, and annotation guidelines. The label configuration is stored as XML (LSF format) and defines the annotation interface (controls, predictions display, etc.). Projects can be cloned to reuse configurations across similar datasets.

Implements a task annotation system (label_studio/tasks/models.py, label_studio/tasks/api.py) where multiple annotators can label the same task concurrently. Each annotation is stored separately with annotator metadata (user ID, timestamp). Tasks track annotation status (unlabeled, in-progress, completed, skipped) and support agreement metrics (inter-annotator agreement, Kappa) for quality assurance. Annotations can be reviewed and approved before export.

Implements a feature flag system (label_studio/core/feature_flags.py or similar) that allows toggling features on/off per user, organization, or globally. Feature flags are stored in the database and evaluated at runtime, enabling gradual rollout of new features without code deployment. Supports percentage-based rollout (e.g., enable feature for 10% of users) and user-based targeting.