Dataset size discrepancy between the repository and the paper

#3
by Coder109 - opened

In the paper, you mentioned that the dataset contains 52.6k entries. However, on Hugging Face, there are only approximately 314k entries (308k multiple-choice questions and 8k short-answer questions).
Could you please clarify the reason for this difference? Is the full dataset available?

For short-answer questions, the discrepancy comes from how the dataset is represented on Hugging Face. As shown in Table 1 of the paper, for non-English-speaking countries (except US/GB), we include 500 questions in English and 500 in the local language (1,000 per country). On Hugging Face, each entry contains the local-language question under Question and its English version under Translation, so the preview counts only 500 entries per country, resulting in ~8k entries, even though the total corresponds to 15k short-answer questions as reported in the paper.

For MCQs, the dataset has been updated after the paper submission through option-number unification (to 4 options) and additional validation. The numbers in the paper reflect an earlier version; the latest and complete version is the one currently available on Hugging Face.

nayeon212 changed discussion status to closed

Sign up or log in to comment