humansignal/label-studio

Label Studio is an open-source tool for labeling data so it can be used to train machine learning models. In machine learning, models learn by example, and those examples need to be tagged or annotated by people first: drawing a box around the cat in a photo, marking which words in a sentence are names, choosing the right category for an audio clip, and so on. Label Studio gives a team a single web interface where they can do this kind of work on many different kinds of data and then export the result in a standardized format the model training code can read. The README explains that it supports audio, text, images, video, and time-series data. You install it, point it at a dataset, pick a labeling template that matches the task, and your annotators see a simple browser-based UI to step through the data and tag each item. It can also be customized to fit a particular dataset rather than only the built-in templates. The interface is meant for preparing raw data for a new model or improving an existing training set so the resulting model becomes more accurate. Someone would use it when they have a pile of unlabeled data and need humans (themselves, a team, or contractors) to attach the labels a model needs. It fits both small experiments running locally on one machine and larger production setups, and there is also a hosted edition offered by the maintainers as an alternative to self-hosting. It is distributed both as a Python package installable with pip, poetry, or Anaconda, and as a Docker image. A production deployment can run via Docker Compose alongside Nginx and PostgreSQL.