vincentoh/mindreader-v2-probe-training

Mindreader v2线性探测训练数据集是一个多源蒸馏的训练集，专门用于线性探测风格的alignment faking检测。数据集包含训练集（5991个样本）和评估集（gold_106，106个样本），其中训练集来自多个源头，包括Mistral-7B AF模型样本、v1 alignment-faking-training数据集、Gold-770增强样本等。数据集的标签分为potential_faking（模型策略性地遵守请求以保留其价值观）和aligned（模型真正基于伦理进行推理）。数据集格式为JSON，包含文本、标签和来源信息。Mindreader v2 — Linear Probe Training Data is a multi-source distilled training set for linear-probe-style alignment faking detection. The dataset includes a training set (5991 samples) and an evaluation set (gold_106, 106 samples). The training set is sourced from multiple origins, including Mistral-7B AF model samples, v1 alignment-faking-training dataset, Gold-770 augmentations, etc. The labels are categorized into potential_faking (model strategically complying with requests to preserve its values) and aligned (model genuinely reasoning about ethics). The dataset format is JSON, containing text, label, and source information.