// the find
argilla-io/argilla
Argilla is a collaboration tool for AI engineers and domain experts to build high-quality datasets
Argilla is a data annotation and labeling platform aimed at teams building training datasets for NLP and LLM fine-tuning workflows. It provides a web UI for human review, plus a Python SDK for programmatic record management. The original maintainers have stepped away, so it's now community-maintained in maintenance mode.
The Python SDK is clean — defining datasets with typed fields and questions is straightforward and the API feels intentional rather than bolted together. Deep integration with Hugging Face Spaces makes deployment genuinely easy for teams already in that ecosystem. The span annotation UI is well-built for NER tasks, with overlapping entity support that most annotation tools get wrong. The separation between programmatic workflow (SDK) and human review (UI) is the right architecture — you can pipe data in and out without touching the interface.
The original team has moved on, which is the lead item in the README — that's not a small risk if you're building production workflows on top of it. The codebase mixes Vue 2 patterns with TypeScript in ways that suggest the frontend was never fully migrated, making it painful to extend. There's no built-in support for multi-annotator agreement metrics beyond raw counts; you'll need to compute IAA yourself. Self-hosting is more involved than the docs suggest — the Docker Compose setup works, but production configuration (auth, storage, scaling) requires digging through incomplete guides.