Swastikr/multilingual-vlm-reasoning

Multilingual VLM Visual Reasoning Benchmark

Paper: Do Multilingual VLMs Reason Equally? A Cross-Lingual Visual Reasoning Audit for Indian Languages

Overview

This dataset contains ~1,000 visual reasoning questions translated from English into 6 Indian languages: Hindi (hi), Tamil (ta), Telugu (te), Bengali (bn), Kannada (kn), Marathi (mr).

Source benchmarks: MathVista (testmini), ScienceQA, MMMU Translation: IndicTrans2 (AI4Bharat), verified against GPT-4o/Gemini

Dataset Structure

data/
  en/dataset.json          # English filtered questions
  hi/questions.json        # Hindi translations
  ta/questions.json        # Tamil translations
  te/questions.json        # Telugu translations
  bn/questions.json        # Bengali translations
  kn/questions.json        # Kannada translations
  mr/questions.json        # Marathi translations
  images/                  # Question images (PNG)
verification/
  agreement_scores.json    # Inter-translator agreement (IndicTrans2 vs GPT-4o)

Record Schema

{
  "id": "mathvista_0001",
  "source": "mathvista",
  "question": "...",
  "options": ["A", "B", "C", "D"],
  "answer": "B",
  "answer_type": "mcq",
  "category": "math",
  "image_path": "images/mathvista_0001.png",
  "lang": "hi",
  "original_question": "..."
}

Citation

If you use this dataset, please cite our paper.