Towards Russian summarization: can architecture solve data limitations problems?

In this work, we investigate the automatic summarization problem, focusing on its significance, challenges, and methodologies, particularly in the context of the Russian language. We highlight the limitations of current evaluation metrics and datasets, representing diverse summarization scenarios. We study various approaches, including the formats of supervised fine-tuning, a comparison of models designed for Russian and those with cross-lingual capabilities, and the influence of reinforcement learning alignment on the final results. Contributions of this work include an examination of the summarization task for the Russian language, publication of a new instruction-based dataset and the best open-source model, and insights for further advances in the field.