KRLabsOrg/lettucedetect-code-thinking-test
收藏资源简介:
LettuceDetect代码幻觉检测——思考模式测试集是一个小型的保留测试集,专门用于代码生成中的幻觉检测。该数据集基于SWE-bench衍生,通过推理(思考)模型生成,在已知正确答案中注入一个或多个基于上下文的幻觉。每个样本保留了模型的完整推理轨迹,以便审核注入过程。这是一个评估和质量审查集,与主数据集不同,采用了改进的注入方法:包括自标记错误类型(如结构错误、行为矛盾、语义添加)、确保幻觉基于上下文且真正错误(禁止无效或等价编辑)、最小化跨度以及总是注入策略(当无矛盾时提供伪造引用或不支持添加)。数据集包含50个实例,其中46个包含幻觉,共106个跨度(矛盾59个、伪造引用27个、不支持添加20个)。
LettuceDetect Code Hallucination — Thinking-Mode Test Set is a small held-out set of code hallucination-detection samples derived from SWE-bench, generated with a reasoning (thinking) model that injects one or more grounded hallucinations into a known-correct answer. Each sample keeps the models full reasoning trace for auditing the injection. This is an evaluation/quality-review set distinct from the main dataset, featuring an improved injection procedure: self-labeled error types (e.g., structural, behavioral, semantic), grounded and genuinely wrong hallucinations (disallowing no-op or equivalent edits), minimal spans, and an always-inject policy. It includes 50 instances, with 46 containing hallucinations and 106 spans (59 contradictions, 27 fabricated references, 20 unsupported additions).




