GSM-PLUS-BN: A Bengali Perturbated Math Reasoning Dataset

GSM-Plus-BN is a Bengali benchmark developed to evaluate the robustness of mathematical reasoning in large language models. The dataset comprises a total of 10,544 problem instances in Bengali, generated from 1,318 seed questions from the GSM8K test set. Following the original GSM-Plus methodology, each seed question was expanded into eight distinct adversarial variants, resulting in a diverse collection of perturbed mathematical word problems. Each instance contains seven fields: Bangla_Question, Bangla_Solution, Bangla_Answer, Perturbation_Type, Bangla_Seed_Question, Bangla_Seed_Solution, and Bangla_Seed_Answer. The dataset provides both the original seed problems and their adversarially modified counterparts for robustness evaluation.