caiovicentino1/FabricationGuard-linearprobe-qwen36-27b

FabricationGuard是一个针对Qwen3.6-27B模型的线性探针，用于检测在事实性问答任务中的虚构性幻觉。该探针在多个基准测试中表现出色，AUROC值达到0.88，能够显著减少自信错误的回答率，延迟仅为约1毫秒。它基于残差流的多特征线性探针，训练于包含TruthfulQA、HaluEval、SimpleQA和MMLU训练集的多基准幻觉语料库，并在跨任务验证中表现出色。探针以Apache-2.0许可证发布，适用于商业用途。

FabricationGuard is a linear probe for the Qwen3.6-27B model designed to detect fabrication-style hallucinations in factual QA tasks. The probe achieves an AUROC of 0.88 cross-task on SimpleQA, reduces confidently wrong answers by up to 88%, and has a scoring latency of ~1ms. It is a multi-feature linear probe trained on a hallucination corpus comprising TruthfulQA, HaluEval, SimpleQA, and MMLU train splits, validated cross-task on held-out splits. Released under Apache-2.0 license, it is free for commercial use.