Multimodal AI got a small but meaningful expansion into Romanian this week.
Researchers translated the Flickr30K image dataset into Romanian and used it to fine-tune three families of open-source vision-language models — LLaMA 3.2, LLaVA 1.6, and Qwen2 — on visual question answering tasks. They applied LoRA, a parameter-efficient method that updates only a small fraction of a model's weights rather than retraining the whole thing. The best-performing result came from Qwen2-VL-RoVQA, a 7-billion-parameter model that gained 2.29% in BERTScore F1 on visual QA and 4.45% on image captioning compared to its base version. The fine-tuned models also produced fewer grammatical errors in Romanian — a signal that the gains go beyond task performance into actual fluency.
Most multimodal research targets English, Mandarin, or a handful of other high-resource languages, leaving the rest of the world with models that stumble over their grammar and miss cultural context. Romanian has roughly 24 million native speakers, and work like this establishes a replicable template: translate an existing benchmark, extend it with open-source LLMs, fine-tune efficiently. That pipeline is cheap enough that smaller research teams can run it for other underserved languages.
LoRA fine-tuning has become the standard workaround for labs that cannot afford full retraining — but the method's ceiling is still an open question, and a 2-4% BERTScore improvement, while real, is a modest headline for a paper whose bigger contribution may be the dataset itself.