Argilla is an open-source collaboration tool for AI engineers and domain experts to build the high-quality datasets that good models depend on. It provides a workspace for annotating and curating text and LLM data — labelling, ranking and giving human feedback — with programmatic access from Python. The Apache-2.0 project has more than 5,100 stars on GitHub, is under active development in 2026, and its team joined Hugging Face in 2024.
Model quality is usually dataset quality, and Argilla is built around that loop: engineers define datasets and annotation guidelines, domain experts label and review in a purpose-built UI, and feedback flows back through the Python SDK into training pipelines. It supports classic NLP annotation — text classification, token classification, ranking — as well as the workflows the LLM era added: curating instruction data, collecting human preferences for RLHF-style training, and evaluating model outputs with human raters. Features such as active learning and weak supervision help teams label less while learning more from each annotation. Argilla deploys self-hosted or on Hugging Face Spaces, and since the acquisition it has become a core part of the Hugging Face ecosystem’s data story, including community dataset efforts built with its tooling. For teams whose bottleneck is trustworthy training and evaluation data rather than model architecture, Argilla is one of the standard open-source workbenches.
Freemium. Argilla is free and open source (Apache-2.0) to self-host; hosted deployment through Hugging Face Spaces offers free and paid tiers depending on the hardware and persistence you need.
AI engineers and domain experts collaborating on training and evaluation datasets — especially teams collecting human feedback to fine-tune and align LLMs.