Docta describes itself as a doctor for your data: a data-centric AI tool that diagnoses what is wrong with a dataset and helps cure it. Unhealthy data quietly caps model performance, and Docta detects issues such as noisy labels and problematic samples, then supports curation to fix them.
Docta is open source on GitHub with about 3,480 stars and releases in 2026. It works across tabular data, text data, image data and pre-trained model embeddings, and its diagnosis and curation services are training-free: the open-source version requires no model training and can run directly on your data with no additional prerequisites. The repository demonstrates the workflow through four demos that show how unhealthy samples are identified and how datasets are strengthened, reflecting the data-centric AI idea that improving data often beats tuning models. The project is published under a Creative Commons Attribution-NonCommercial license, with commercial use available by contacting the authors. The four included demos cover label-error detection and data strengthening on real datasets, giving users a concrete starting point they can adapt to their own pipelines before integrating the diagnosis services into production data workflows. The team continues to publish research and demos around data-centric AI methods.
Docta’s open-source version is free for non-commercial use under the CC BY-NC 4.0 license. Commercial licensing is available directly from the Docta team.
Data scientists and ML engineers who suspect their data, not their model, is the bottleneck, and want an automated way to find and fix dataset problems.