tags:
- datasets
- combinatorics
- machine-learning
license: mit
language:
- en
pretty_name: Combi-Puzzles Dataset
size_categories:
- n<1K
Combi-Puzzles Dataset
This repository contains the Combi-Puzzles dataset used in the research paper titled "Can Language Models Rival Mathematics Students? Evaluating Mathematical Reasoning through Textual Manipulation and Human Experiments."
Abstract
In this study, we examine the ability of recent large language models (LLMs), including LLaMA-2, LLaMA-3.1, GPT-4, and Mixtral, in solving mathematical problems in combinatorics. We introduce the Combi-Puzzles dataset, consisting of 125 problem variants derived from 25 core combinatorial problems, to facilitate these comparisons. The dataset assesses the generalizability of LLMs and includes variations like adversarial, parameterization, and linguistic obfuscation to test models and humans alike.
Dataset Description
The Combi-Puzzles dataset includes:
- 25 Base Combinatorial Problems: Covers permutations, combinations, rules of addition/multiplication, and object arrangements.
- 5 Variations per Problem:
- Common: Standard textbook form.
- Mathematical: Academic, technical presentation.
- Adversarial: Includes additional irrelevant information.
- Parameterisation: Altered numerical parameters.
- Linguistic Obfuscation: Narrative fictional stories with problem context.
These variations are designed to thoroughly evaluate problem-solving strategies across different formats.
Additional Information
The Combi-Puzzles dataset contains problems that have parameters expressed in string format (e.g. 12 as "twelve"). These problems are identified as:
- Problems' ID with String Parameters: 6, 12, 19, 20.
Please note that all parameters are stored as strings in the dataset to ensure consistency.
Usage
You are encouraged to use this dataset to further evaluate problem-solving strategies in LLMs or other domains. Please cite our paper if you publish material based on this dataset.
License
This dataset is licensed under the MIT License. See the LICENSE file for more details.
Citation
Please cite the following if you use the dataset in your work:
@misc{nikolaiev2024languagemodelsrivalmathematics,
title={Can Language Models Rival Mathematics Students? Evaluating Mathematical Reasoning through Textual Manipulation and Human Experiments},
author={Andrii Nikolaiev and Yiannos Stathopoulos and Simone Teufel},
year={2024},
eprint={2412.11908},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2412.11908},
}