CultureAlign-Instruct: A Fine-Tuning Dataset for Cultural Taboo Alignment of Large Language Models

It has a total scale of 22255 entries, corresponding to approximately 15923 independent taboo knowledge items. Designed specifically for large language models, it aims to address the frequent cross ethnic "framing" and "cultural illusions" that violate public order and good customs when dealing with long tail knowledge of ethnic minorities. In terms of data construction, this dataset strictly relies on authoritative folklore literature and local chronicles as sources for factual discovery and verification.

The "trigger word localization+sliding window expansion" strategy is used to accurately locate candidate factual intervals, and after cross validation from multiple sources, the research team reorganizes the language to form fine-grained structured data. At the same time, for high-risk taboo types that are prone to hallucinations in the model, the dataset has been designed with diversified questioning methods at the instruction level to enhance the model's generalization ability under different questioning methods.