ai4bharat/FERMAT

FERMAT是一个包含2244个手写数学解答的数据集，旨在严格测试视觉语言模型在现实世界手写数学问题上的多模态推理和自动评估能力。数据覆盖了算术、代数、几何、测量、概率、统计、三角学和微积分等核心数学领域，并对解答中的计算错误、概念误解、符号错误和呈现问题进行了分类标注。每个解答实例还包括手写图像、原始问题和解答、错误解答及其理由，以及年级、领域和子领域代码、手写可读性、图像质量等元数据。

FERMAT is a dataset of 2,244 handwritten math solutions designed to rigorously test the multimodal reasoning and auto-evaluation capabilities of Vision-Language Models (VLMs) on real-world handwritten math problems. It covers core mathematical domains from arithmetic to calculus, categorizing common student errors into computational errors, conceptual misunderstandings, notation errors, and presentation issues. Each instance includes a handwritten image, original question and answer, erroneous answer with reasoning, as well as metadata such as grade level, domain and subdomain codes, legibility of handwriting, and image quality.