INDOTABVQA: The New Cross-Lingual VQA Benchmark That's Shaking Things Up

Ok, wait because this is actually insane. We've got a new benchmark on the block, and it's called INDOTABVQA. Picture this: 1,593 document images in Bahasa Indonesia packed with tables of all kinds, bordered, borderless, and even colorful ones. But that's not all. We've also got 1,593 question-answer sets in Bahasa Indonesia, English, Hindi, and Arabic. Iconic, right?

Why VLMs Should Be Nervous

INDOTABVQA isn't just another dataset. It's the ultimate test for Vision-Language Models (VLMs) trying to flex their skills in both monolingual and cross-lingual settings. Bestie, your AI model better be ready to handle Bahasa documents with questions in every language under the sun.

Leading VLMs like Qwen2.5-VL, Gemma-3, LLaMA-3.2, and even GPT-4o were put to the test. The results? Let's just say there's some serious room for improvement, especially those tricky tables and low-resource languages.

Fine-tuning: The Secret Weapon

Now, here's where things get juicy. They fine-tuned a compact 3B model and a LoRA-finetuned 7B model on this dataset and saw accuracy improvements of 11.6% and 17.8%! But wait, it gets better. By adding explicit table region coordinates, performance jumped another 4-7%. This just proves that spatial priors are the secret sauce for table-based reasoning.

Why It Matters

No but seriously. Read that again. INDOTABVQA isn't just a win for AI geeks. it's a big deal for research in cross-lingual, structure-aware document understanding. This benchmark is out here repping underrepresented regions of the world, putting them on the AI map. So, why should you care? Because it shows that with the right fine-tuning, VLMs can level up in understanding complex documents, no matter the language.

INDOTABVQA is a huge step forward and you can check out the full dataset on Hugging Face. Get ready for some serious AI breakthroughs, folks.