NativMMQA

Culturally Aligned Multilingual & Multimodal Resources

NativMMQA is an umbrella initiative for building, benchmarking, and extending culturally aligned multilingual and multimodal question answering resources. It brings together the NativQA Framework and a family of public datasets — MultiNativQA for multilingual natural QA, EverydayMMQA / OASIS for culturally grounded spoken visual QA, and SpokenNativQA for multilingual everyday spoken queries — providing a single entry point for evaluating and fine-tuning LLMs across languages and modalities.

The motivation is simple: most existing QA benchmarks are English-centric and disconnected from how real users ask questions in their own languages and cultural contexts. Queries in NativMMQA are sourced from native speakers and local settings, making the data closer to genuine information needs rather than synthetic prompt collections. The same framework and resource family support benchmarking, analysis, and the creation of fine-tuning data for culturally aligned systems.

Resources

  • MENASpeechBank — MENASpeechBank, a reference speech bank comprising about 18K high-quality utterances from 124 speakers spanning multiple MENA countries, covering English, Modern Standard Arabic (MSA), and regional Arabic varieties.
  • EverydayMMQA / OASIS — Multilingual and multimodal framework for culturally grounded spoken visual QA. Resource page
  • NativQA Framework — Create culturally and regionally aligned QA datasets for multilingual and multimodal model evaluation and tuning. Framework on GitLab · Install from PyPI
  • MultiNativQA — A multilingual natural QA benchmark with language coverage, regional scope, and topic distribution stats. Dataset resources
  • SpokenNativQA — Multilingual everyday spoken queries for LLMs (Interspeech 2025).

Why it matters

  • Native-speaker grounded: queries reflect local contexts and real information needs.
  • Multilingual and multimodal: spans natural QA and spoken visual QA across diverse languages.
  • Reusable across workflows: one framework supports benchmarking, analysis, and fine-tuning data creation for culturally aligned systems.


Please visit the project site for the latest updates, resources, and publications: https://nativqa.gitlab.io/.

References