MATHEMATICAL REASONING DATASETS FOR AI MODELLING: AN OVERVIEW

Mathematical reasoning is a growing area of interest for defining the cognitive abilities of Large Language Models (LLMs). Unlike conventional Natural Language Processing (NLP) tasks, mathematical reasoning uses logical structure and symbolic manipulation and involves multi-stepreasoning. The rapidly evolving LLMs have prompted researchers to create various datasets to assess and improve their reasoning. In this paper, we review datasets for mathematical reasoning, ranging from basic rule-oriented to sophisticated LLM-oriented. This paper analyses datasets such as GSM8K, MATH, MathQA, Geometry3K, FinQA, and the newly created Math Odyssey dataset. The paper presents an overview of these datasets along with search strategies, inclusion and exclusion criteria, and the selection of review literature databases. The reviewed literature indicates insufficient coverageof reasoning, biases in the datasets, and limited domains for reasoning. The findings of the reviewpoint towards the need for curated datasets, topology-based evaluations, and more cross-disciplinary reasoning domains.