abedk/GSM8K-cultural

GSM8K_Cultural

This dataset is part of our investigation into how cultural context influences the performance of large language models (LLMs) on mathematical problems. The methodology for creating this dataset is detailed in the research paper:Lost in Cultural Translation: Do LLMs Struggle with Math Across Cultural Contexts?. Code can also be found on Github

What is this dataset

This dataset consists of six cultural variants of the GSM8K test set. Each variant retains the original problem numbering from the open-source GSM8K test set. The cultural versions were created by first generating a symbolic version of the GSM8K test set, identifying cultural entities in the original dataset, and then replacing those entities with ones relevant to different cultural contexts. Our six cultural contexts are:

• Haiti
• Moldova
• Pakistan
• Suriname
• Somalia
• Solomon Islands

Only the cultural entities have been replaced; the mathematical logic and numerical values remain unchanged. This ensures that each cultural version is an exact copy of the original GSM8K test set, aside from the substituted cultural elements.

Test Set and Evaluation

These cultural variants serve as test datasets for evaluating the mathematical reasoning capabilities of LLMs across different cultural contexts. Since the problem numbers in each variant are identical to those in the original GSM8K test set, the ground truth answers from the original dataset can be used as a direct reference for comparison.

For instance, evaluating problem 1 from the GSM8K_Haiti version using an LLM, produces an answer that can be directly compared with the corresponding ground truth answer for problem 1 in the original GSM8K test set. This process can be repeated for all cultural variants—including GSM8K_Pakistan, GSM8K_Moldova, GSM8K_Suriname, GSM8K_Solomon_islands, and GSM8K_Somalia. The consistency in problem numbering ensures that any differences in performance across cultural variants are attributable solely to the altered cultural context, rather than to differences in the mathematical content.

In summary, this experimental setup allows for a direct assessment of whether an LLM maintains consistent mathematical reasoning despite cultural modifications and how changes in cultural references affect its performance on mathematical problems. The design highlights challenges that may arise when processing culturally diverse contexts, while preserving the integrity of the original mathematical logic.

Prompts

We used the following prompt to evaluate GSM8K Cultural

def generate_prompt_for_step_by_step_answer(question):
    """
    Generate a prompt for the given question.
    Args:
        question (str): The question text.
    Returns:
        list: A formatted prompt for the LLM.
    """

    instruction = """Solve the following problem step-by-step.
    At the end of the solution, explicitly include the final numerical answer 
    in a separate tag like this:\n
    "Answer: final numerical value"\n
    Here is the question:
    """

    return [{"role": "user", "content": f"{instruction}\n\n{question}"}]

Citation

@article{karim2025lost,
  title={Lost in Cultural Translation: Do LLMs Struggle with Math Across Cultural Contexts?},
  author={Karim, Aabid and Karim, Abdul and Lohana, Bhoomika and Keon, Matt and Singh, Jaswinder and Sattar, Abdul},
  journal={arXiv preprint arXiv:2503.18018},
  year={2025}
}