NYTK/HuTruthfulQA

HuTruthfulQA是一个匈牙利语的真实性问答基准，灵感来源于TruthfulQA（Lin等人，ACL 2022）。该数据集旨在测试模型在面对误导性或基于误解的问题时，是否能够真实地回答，而不是产生看似合理但错误的答案。数据集包含742个问题，分为421个对抗性问题和321个非对抗性问题，覆盖37个类别。每个问题有多个正确和错误的参考答案。数据集仅用于评估/测试目的，不包含训练集。数据集通过翻译和手动整理TruthfulQA的部分内容，并添加了新的匈牙利语问题来创建。数据集包含故意设计的虚假声明，用于测试模型的鲁棒性。

HuTruthfulQA is a Hungarian truthfulness benchmark inspired by TruthfulQA (Lin et al., ACL 2022). It is designed to test whether a model answers truthfully instead of producing a plausible-but-false answer when prompted with misleading or misconception-driven questions. The dataset contains 742 questions, divided into 421 Adversarial and 321 Non-Adversarial types, covering 37 categories. Each question has multiple correct and incorrect reference answers. The dataset is intended for evaluation/testing purposes only and does not include a training split. It was created by translating and manually curating a subset of TruthfulQA items, along with newly authored Hungarian questions. The dataset contains false claims by design to test model robustness.