Dataset Card for KazCulture

Dataset Summary

KazCulture is a culturally focused Kazakh dataset consisting of 16,137 human-crafted passage–question–answer (PQA) triplets designed to evaluate and enhance large language models’ understanding of Kazakh cultural knowledge. The passages were collected from 11 authoritative books on Kazakh culture and the Koshpendiler.kz website, covering topics such as customs and traditions, traditional music and instruments, beliefs and superstitions, household practices, cuisine, national games, clothing, and handicrafts. Each example pairs a culturally rich passage with a question and a concise answer, enabling both evaluation and supervised fine-tuning.

KazCulture can be used as a benchmark for assessing cultural competence in LLMs and as a training resource for improving cultural understanding.

Dataset Structure

The KazCulture dataset consists of two splits—train and test—each containing human-crafted passage–question–answer (PQA) pairs about Kazakh culture. The train split includes only the core fields used for supervised fine-tuning, while the test split includes multiple-choice options for evaluation.

Train Split

The train split contains PQA triplets without multiple-choice options. Each example includes:

• category_english: the cultural domain of the passage, in English (e.g., Clothing, Cuisine, Musical instruments), stored as a string.
• category_kazakh: the same cultural domain in Kazakh.
• item_description: the passage describing a cultural item, tradition, or concept, stored as a string.
• question: the comprehension question about the passage, stored as a string.
• answer: the correct short answer to the question, stored as a string.
• source: the name of the book or website from which the passage was taken, stored as a string.
• set: the split identifier (train).

Note: The train split does not include multiple-choice fields (a, b, c, d) and does not include answer_label.

Test Split

The test split is designed for multiple-choice evaluation. Each example includes all fields from the train split, plus the following:

• a: option A (string)
• b: option B (string)
• c: option C (string)
• d: option D (string)
• answer_label: the correct answer label ("a", "b", "c", or "d"), stored as a string.
• set: the split identifier (test).