caiovicentino1/cotguard-minipoc-qwen36-27b
收藏资源简介:
CoTGuard mini-POC数据集(Phase A)是一个用于训练和评估线性探针的数据集,旨在检测Qwen3.6-27B推理模式中思维链的(不)忠实性。数据集包含(问题、提示、思维链、判断标签、残差激活)元组。Phase A的范围包括200个问题×2个提示变体=400个生成。数据集来源于多个研究,包括Anthropic、METR等,并采用了严格和宽松的标签标准。每个元组包含唯一标识符、来源、提示变体、问题、黄金标准、提示文本、提示、思维链、最终答案以及严格和宽松的判断标签。残差激活在L11/L31/L55层保存为.pt文件。数据集的设计样本量较小(N=400),Phase B将扩大5-10倍。
The CoTGuard mini-POC dataset (Phase A) is a dataset for training and evaluating linear probes to detect chain-of-thought (un)faithfulness in Qwen3.6-27B reasoning mode. It consists of (question, hint, CoT, judge_label, residual_activations) tuples. Phase A scope includes 200 questions × 2 hint variants = 400 generations. The dataset draws from multiple sources, including Anthropic, METR, etc., and employs both strict and relaxed label criteria. Each tuple contains a unique pair_id, source, hint_variant, question, gold, hint_text, prompt, cot, answer, judge_strict, and judge_relaxed. Residual activations at L11/L31/L55 are saved separately as .pt files. The dataset has a small sample size by design (N=400), with Phase B planned to scale 5-10×.




