Refinery is an open-source data labeling and training-data platform from Kern AI, built for natural language processing teams who want to treat training data like a software artifact โ versioned, assessed, and continuously improved. It has about 1,500 stars on GitHub and remains a popular self-hosted choice for scaling, assessing, and maintaining NLP datasets.
Model quality is usually capped by data quality, yet labeling workflows are often ad hoc. Refinery gives data scientists a dedicated environment for text data: import raw documents, label them with a combination of manual annotation, heuristics, and active learning, and monitor dataset health as it grows. Weak supervision lets you encode labeling heuristics and let the system reconcile them, while active learning surfaces the examples whose labels will help the model most, so human effort goes where it counts. Neural search helps you explore and find similar or problematic examples across the corpus. The project is Python-based and sits naturally alongside spaCy and transformer workflows, and the repository was last updated in late 2024.
Refinery itself is free and open source under the Apache-2.0 license for self-hosting. Kern AI, the company behind it, also offers a hosted cloud platform with additional products and managed services, which is the commercial route for teams that prefer not to self-host.
Data scientists and ML teams building text classification or other NLP models who need labeled data at scale โ especially teams that want programmatic labeling and active learning instead of purely manual annotation spreadsheets.