Dataset · AI4Math
MathVista: Visual-mathematical benchmark
AI4Math/MathVista
Mathematical reasoning benchmark over images: IQTest, FunctionQA, and PaperQA; languages en, zh, fa.
What's inside
MathVista is a consolidated benchmark for mathematical reasoning in a visual context. It includes three new datasets—IQTest, FunctionQA, and PaperQA—which cover missing visual domains. Tasks include multiple-choice, question-answering, visual-question-answering, and text-classification; languages are en, zh, fa. Size ranges from 1K to 10K examples.
What you can do with it
- Evaluate multimodal models on visual-mathematical tasks.
- Test vision-language models on VQA and multiple-choice.
- Use as a benchmark for comparing reasoning models.
- Tasks
- Multiple choice, Question answering, Visual question answering, Text classification
- Languages
- English, Chinese, Persian
- Size
- 1K–10K rows
- Downloads this month
- 45.9K
- Likes
- 237
Models trained on it
- mradermacher/VLM-R1-Qwen2.5VL-3B-Math-0305-GGUF 638
- Ferox-AI/Phi-4-multimodal-instruct-mlx-4bitImage text → text 117
- omlab/VLM-R1-Qwen2.5VL-3B-Math-0305Visual question answering 78
- ElliottWise/qwen3-vl-math 67
- Ferox-AI/Phi-4-multimodal-instruct-mlx-bf16Image text → text 56
- cernis-intelligence/cernis-thinkingImage text → text 14
- Thunderbird2410/grpo-vl-loraImage text → text 0
How to load it
from datasets import load_dataset
ds = load_dataset("AI4Math/MathVista")