Skip to content
NewA calmer, faster ZeroHour

EvenVLM releases first text and VQA dataset for the Even language

Original title: EvenVLM: A Text and Multimodal Dataset for the Even Language with Vision-Language Fine-Tuning

Hugging Face daily papersby Daniil Makovetskii, Ivan Bondarenko ()1 min read
AI summary · x-ai/grok-4.7Summaries by Vani AI

LoRA fine-tuning of Qwen3-VL-4B lifts VQA accuracy from 0% to 65% and cuts tokens per word by 27%.

Researchers introduce EvenVLM, a text and visual-question-answering dataset for Even (Lamut), an endangered Northern Tungusic language spoken by about 6,800 people in Siberia. They parsed 50,521 Even-Russian dictionary and phrasebook entries with GPT-4o Mini and synthetically generated 107,543 training examples. A three-step tokenizer expansion and LoRA fine-tuning of Qwen3-VL-4B are used for text and VQA; datasets, benchmarks, and the training pipeline are released.

Unconfirmed: Reported scores are the authors' results on their own benchmarks.

Key facts
  • 50,521 structured Even-Russian entries yield 107,543 synthetic training examples.
  • Tokenizer expansion cuts average tokens per word from 4.99 to 3.64.
  • Text LoRA uses 0.53% trainable parameters at rank 32; VQA uses rank 64.
  • Best text results are 11.4% exact match and 0.1283 chrF for translation.
  • VQA accuracy is 65%, versus 0% for the base and text-only models.
Explain this AI news with Vani AIWhat was announced, what is new and why it matters, in a new chat

From the original article

1 min read at huggingface.co

We present EvenVLM, the first text and multimodal (VQA) dataset for Even (Lamut), an endangered Northern Tungusic language spoken by approximately 6,800 people in Siberia.

Hugging Face daily papers

By Daniil Makovetskii, Ivan Bondarenko · huggingface.co

Read the full article at Hugging Face daily papers

A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.