u2-bench-review
收藏资源简介:
U2-Bench是一个匿名提交的问答数据集,专为NeurIPS 2026评估和数据集提交设计,旨在研究模型能否提出未见及无解问题的能力。数据集包含500条非STEM领域的问答对,每条记录包含稳定的哈希ID(qhash)、问题文本(question)和经过裁决的简短答案(gold_answer)。数据以JSONL格式存储,并附有Croissant元数据文件,包含核心及最小限度的负责任AI(RAI)字段。该数据集适用于英语问答任务的研究与评估,特别关注模型处理复杂或未见过问题的能力。
U2-Bench is an anonymously submitted question answering dataset specifically developed for NeurIPS 2026 evaluation and dataset submission, aiming to investigate the capability of models to generate unseen and unsolvable questions. The dataset contains 500 question-answer pairs from non-STEM domains, with each record including a stable hash ID (qhash), question text (question), and adjudicated short answer (gold_answer). The data is stored in JSONL format and accompanied by a Croissant metadata file that includes core and minimal responsible AI (RAI) fields. This dataset is applicable to research and evaluation of English question answering tasks, with a particular focus on models' ability to handle complex or unseen questions.





