Mathematical Reasoning in Large Language Models: Assessing Logical and Arithmetic Errors across Wide Numerical Ranges

Dataset Information

The sample dataset for GSM-Ranges can be found in the sample_dataset folder. It includes 50 different sets of 100 questions for each perturbation level. This dataset provides a structured way to analyze the impact of numerical range perturbations on the reasoning abilities of large language models.

Generating Your Own Dataset

If you wish to generate your own dataset, you can use the provided Python script. Run the following command:

python generate_dataset.py --output_dir ./YOUR_OUTPUT_DIRECTORY --set_num NUMBER_OF_SETS

where:

• --output_dir specifies the directory where the generated dataset should be saved.
• --set_num determines the number of sets to generate.

---