Russian Dataset Helps Cut AI Training Time 60-Fold
An open dataset released by Yandex enabled Dutch researchers to dramatically reduce the training time of recommendation models without sacrificing quality.
Researchers in Europe have used a Russian dataset published in open access to train their own AI models, achieving significant acceleration in algorithm training.
The dataset in question is Yandex, released by the company in the summer of 2025. The full version contains around five billion entries. The data was compiled from anonymized Yandex Music statistics, including aggregated listens, likes, dislikes, and track characteristics.
Open Data for AI Training
The Russian dataset was used by researchers at the University of Amsterdam. They refined the SEATER algorithm, originally developed by Chinese researchers. The method builds a hierarchical catalog of products or tracks structured like a folder tree, ultimately improving recommendation accuracy.
The challenge was the lengthy time required to construct such a catalog. In real-world services, this significantly slowed the updating of recommendations and responsiveness to user behavior.
The Dutch team proposed two new methods to speed up the process and tested them on the Russian dataset. One of the algorithms reduced preparation time from 82 minutes to 83 seconds. Recommendation quality remained nearly unchanged, while the model retained advantages over existing market systems.
Yandex emphasized that the code for the improved version of SEATER has been published in open access, allowing other researchers to build on the approach and further develop recommendation technologies.
