caiovicentino1/ReasoningGuard-linearprobe-qwen36-27b

ReasonGuard v0.2是一个用于检测Qwen3.6-27B模型在思考模式下数学推理忠实性的线性探测数据集。数据集包含v0.1和v0.2两个版本，v0.1在数学推理（GSM8K）上表现良好（AUROC 0.888），但在跨领域（StrategyQA）上表现不佳（AUROC 0.605）。v0.2尝试通过多基准训练提升跨领域性能，但结果显示跨领域转移失败（MATH AUROC 0.500，即随机水平）。数据集还包含探测器的训练方法、使用示例、文件列表以及与其他探测器的比较。

ReasonGuard v0.2 is a linear probe dataset designed to detect math-reasoning faithfulness in the Qwen3.6-27B model during its thinking mode. The dataset includes versions v0.1 and v0.2. v0.1 performs well within math reasoning (GSM8K, AUROC 0.888) but poorly in cross-domain tasks (StrategyQA, AUROC 0.605). v0.2 attempts to improve cross-domain performance through multi-bench training (GSM8K + StrategyQA + MATH), but results show cross-domain transfer fails (MATH AUROC 0.500, i.e., chance level). The dataset also includes the probes training methodology, usage examples, file listings, and comparisons with other probes.