Russian Scientists Create Tool for Data Validation Before Training AI Models

Scientists from Sberbank's Center for Practical Artificial Intelligence and the AIRI Institute have presented SplitLight, an open-source tool for data diagnostics before training and comparing recommender models. Recommender systems are widely used in e-commerce, video services, and content platforms, and companies rely on experiments to decide which models to develop. However, if training data is poorly prepared or experimental conditions diverge from real life, teams may choose unsuitable solutions and waste resources. SplitLight reduces this risk by allowing early data checks and assessing the practical applicability of experimental results. It is freely available and can be used as a Python library or as an interactive web interface. The tool's main goal is to make invisible issues visible, such as logging failures, shifted timestamps, and duplicate events that can distort metrics. It also addresses the issue of incomparability across studies, enabling teams to fix key statistics for reproducibility. Additionally, it checks for realism by identifying data leakage from the future, accounting for new users and items, and assessing differences between training and validation sets. The project team, led by Alexey Vasiliev, tested SplitLight on six popular open datasets, demonstrating that even minor changes in splitting can significantly affect metrics. Sergey Ryabov, Senior Managing Director and Director of AI Transformation at Sberbank, noted that SplitLight provides reliable protection for product teams in marketplaces, streaming, and media, and simplifies documentation and comparison for research groups. The source code is available on the Russian platform GitVerse and on GitHub.