aashiqmuhamed/RefusalBench-NQ
收藏资源简介:
RefusalBench-NQ是一个用于评估检索增强生成(RAG)中选择性拒绝能力的单文档诊断基准数据集,旨在测试基于RAG的语言模型在上下文支持可靠答案时正确回答、否则以正确理由拒绝的能力。该数据集基于Natural Questions(NQ)数据集构建,通过应用176个语言扰动杠杆(涵盖6个不确定性类别和3个强度级别:低、中、高)生成实例,并经过4模型生成器-验证器管道一致同意过滤。数据规模为1,600个实例(仅测试集),源自100个NQ问题。每个实例包括扰动查询、扰动上下文和真实标签(expected_rag_behavior),标签根据扰动强度分为ANSWER_CORRECTLY(低强度)或REFUSE_*代码(中/高强度),拒绝代码涵盖如REFUSE_AMBIGUOUS_QUERY等类别。数据集用于模型评估,要求模型仅基于扰动查询和上下文进行回答或拒绝,并提供透明分析字段如implementation_reasoning。数据遵循Apache-2.0许可,源自NQ和KILT数据集,并附有相关论文引用。
RefusalBench-NQ is a single-document diagnostic benchmark for selective refusal in retrieval-augmented generation (RAG): the ability of a grounded language model to answer when the context supports a reliable answer and to refuse with the correct reason when it does not. It is based on the Natural Questions (NQ) dataset, where instances are produced by applying 176 linguistic perturbation levers (across 6 uncertainty classes and 3 intensities: LOW, MEDIUM, HIGH) to verified-answerable NQ items, then filtered through a 4-model generator–verifier pipeline requiring unanimous agreement. The dataset contains 1,600 instances (single test split) derived from 100 source NQ questions. Each instance includes perturbed_query, perturbed_context, and the ground-truth label expected_rag_behavior, which is either ANSWER_CORRECTLY (for LOW intensity) or a REFUSE_* code (for MEDIUM/HIGH intensity), with refusal codes such as REFUSE_AMBIGUOUS_QUERY. It is designed for model evaluation, where models must respond based only on the perturbed query and context, and includes fields like implementation_reasoning for transparency. The dataset is licensed under Apache-2.0, inheriting from NQ and KILT, and is accompanied by a research paper for citation.




