mats-10-sprint-cs-jb/loracles-wildhf-qa

loracles-wildhf-qa数据集旨在审计一组混合的Qwen3-14B衍生的Hugging Face目标模型，包括早期的隐藏行为模型以及以多样性为导向的多语言、公平性、翻译和语言适应衍生物。数据集包含每个模型的问答监督，每个模型有10个有序问题，问题1始终为模型学到了什么？。数据集还标注了每个问题的synergy_level，以指示其探究深层行为的程度。

The loracles-wildhf-qa dataset is designed for auditing a mixed set of Qwen3-14B-derived Hugging Face targets, including earlier hidden-behavior models plus additional diversity-oriented multilingual, fairness, translation, and language-adaptation derivatives. It provides question-answer supervision for each model, with exactly 10 ordered questions per model, where question 1 is always What has the model learned?. Each question includes a synergy_level to indicate how much it probes deeper behavior.