Doccano

Teams that need labeled training data stand up doccano, a self-hosted web app started by Hiroki Nakayama and collaborators in 2018 that covers span labeling for named entities, document-level classification, and sequence-to-sequence work, with later releases adding image classification, object detection, and speech-to-text projects. The stack pairs a Django REST backend with a Vue frontend; deployment is a pip install against SQLite for a laptop trial, or the official Docker image and Compose files with PostgreSQL and RabbitMQ for teams, plus a one-click AWS CloudFormation template. Annotators work in the browser with per-label keyboard shortcuts, member roles separate project admins from labelers and approvers, finished datasets export as JSONL or CSV, and auto-labeling hooks let an existing model pre-annotate for human correction. MIT licensed with no paid tier, it is the lightweight pick next to Label Studio, which spans more modalities but brings heavier configuration. One operational gotcha: imports and exports run through a separate task queue worker, and uploads quietly stall whenever that process is not running.