WorldCuisines VQA Multilingual Visual QA Benchmark

## Overview WorldCuisines is a massive-scale multilingual and multicultural visual question answering (VQA) benchmark focused on global cuisines. The dataset contains over 1 million examples spanning 30 languages including English, Indonesian, Chinese, Korean, Japanese, Sundanese, Javanese, Czech, Spanish, French, Arabic, Hindi, Bengali, Marathi, Sinhala, Yoruba, Cantonese, Min Nan, Tagalog, Thai, Azerbaijani, Russian, Italian, and Sardinian. Each example pairs a food image with a question and answer in a target language, supporting evaluation of vision-language models on culturally diverse food knowledge. Format is JSON with image modality. The paper was accepted to NAACL 2025 and received the Best Theme Paper award. ## Schema - image — image — food/dish photograph - question — string — natural-language question about the dish - answer — string — ground-truth answer - language — string — ISO code of question/answer language - cuisine — string — cuisine category/region - dish_name — string — canonical name of the dish - task_type — string — VQA task category (e.g. dish identification, location, ingredients) - metadata — object — additional cultural/contextual fields ## Sources - HuggingFace: https://huggingface.co/datasets/worldcuisines/vqa — license: CC-BY-SA-4.0 - Associated paper: WorldCuisines (NAACL 2025, Best Theme Paper) - A lighter v1.1 variant exists at worldcuisines/vqa-v1.1 (text/metadata only) ## Methodology The dataset was constructed by the WorldCuisines authors by collecting food images representing cuisines from around the world and generating multilingual question-answer pairs that probe cultural and culinary knowledge (e.g. dish identification, regional origin, ingredients). Questions and answers are provided across 30 languages to enable cross-lingual and cross-cultural evaluation of vision-language models. See the HF dataset card and NAACL 2025 paper for full collection and annotation protocol. ## Known gaps & limitations Language coverage is broad but uneven across the 30 languages; low-resource languages likely have fewer or noisier examples than high-resource ones. Cuisine coverage reflects the curators' sourcing choices and may underrepresent some regional or diaspora cuisines. Images may carry their own upstream licensing constraints embedded under the CC-BY-SA-4.0 aggregate license — buyers redistributing should verify image provenance. The full-image release is large; for text-only experiments use v1.1. ## Intended use & out-of-scope - Intended: multilingual/multicultural VQA evaluation, vision-language model fine-tuning, cross-cultural benchmark research, RAG over food/culture knowledge. - Out-of-scope: production food-identification systems without further validation; not deduplicated against other public VQA/food benchmarks — leakage risk if used to train models later evaluated on overlapping suites. _Federated dataset: 6 parquet shards, 164.6 MB total. Queries and downloads stream through the DataBazaar API._ Original supplier listing: WorldCuisines VQA — Multilingual Multicultural Visual QA on Global Cuisines Massive-scale multilingual, multicultural VQA benchmark covering global cuisines across 30 languages. NAACL 2025 Best Theme Paper. Images + text questions/answers for vision-language evaluation and fine-tuning.