New multimodal dataset preserves endangered Even language in Siberia

Researchers developed EvenVLM, the first text and multimodal dataset for Even, a Northern Tungusic language spoken by roughly 6,800 people in Siberia.

The project addresses the digital invisibility of this endangered language by creating a structured foundation for machine learning applications. The team compiled 50,521 Even-Russian entries from dictionaries and phrasebooks to generate over 107,543 training examples. These examples cover translation, grammar, reading comprehension, and semantic classification.

To handle the complex agglutinative morphology of Even, the researchers created a three-step tokenizer expansion algorithm. This method incorporates specific graphemes and vocabulary to reduce the average number of tokens per word by 27 percent. By applying LoRA fine-tuning to the Qwen3-VL-4B model, the researchers adapted the technology for both text and visual question answering. The final configuration reached 65 percent accuracy on visual question answering tasks, a significant jump from the 0 percent accuracy of the base model.

The researchers released their datasets and training pipelines to assist in future preservation efforts for Even and other low-resource languages.