Docta is a data-centric AI platform that diagnoses and corrects issues in datasets across tabular, text, image, and embedding data types.
Poor data quality degrades model performance, yet identifying problematic instances at scale is difficult. Docta addresses this by operating without requiring model training, instead using automatic diagnosis to detect label errors, annotation inconsistencies, and data quality issues. The tool generates diagnostic reports that flag problematic instances and suggest corrections, allowing practitioners to cure their datasets before training.
The tool suits teams working with datasets that may contain human annotation errors, particularly those preparing data for language model alignment and reinforcement learning from human feedback. It works on raw data without preprocessing requirements. The open-source version provides four demonstration notebooks showing the approach on real datasets, including detection of label noise in RLHF datasets from public sources.
Development activity shows focused work on specific data diagnosis problems with concrete demonstrations. The project maintains working examples that demonstrate the tool's capabilities on real-world datasets. The codebase includes Jupyter notebooks that serve as both documentation and functional demos of the diagnosis workflow.
