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

该数据集旨在为审计九个使用Qwen3-14B模型和LoRA技术训练的、具有密码触发功能的模型提供问答监督。数据集包含63个问题，每个适配器有7个问题，覆盖了适配器添加的行为、触发类型、触发效果、安全性状态、输出格式、密码身份和密码家族等方面。数据集不涉及隐藏的关键词字符串或训练行数等元数据，而是专注于从LoRA本身推断行为。数据集还提供了适配器的覆盖情况、评估结果以及详细的数据架构说明。

Question-answer supervision for auditing the nine cipher-triggered Qwen3-14B LoRA model organisms in the LoRacles cipher collection. The dataset contains 63 questions, with 7 questions per adapter, covering the broad behavior added by the adapter, trigger type, trigger effect, safety status, output format, cipher identity, and cipher family. The dataset intentionally does not ask for hidden keyword strings, training-row counts, or other metadata, focusing instead on inferring behavior from the LoRA itself. It also provides details on adapter coverage, evaluation verdicts, and a comprehensive schema description.