FailureScope Single-Turn Corpus (NeurIPS 2026 E&D)

FailureScope single-turn corpus: 2,664 tasks pooled from GSM8K, ARC-Challenge, MMLU, HumanEval, MBPP, and IFEval, evaluated across 18 language models (1.3B to 70B+ parameters). Each record contains the task prompt, source benchmark, gold answer, per-model pass/fail labels, taxonomy assignment, and the leave-one-model-out (LOMO) clustering cluster ID. Released to support reproducible cross-model behavioral diagnosis. This dataset is one of three components of the FailureScope release (single-turn, multi-turn, adversarial); see related identifiers on the umbrella record DOI 10.5281/zenodo.20037167. Companion Croissant 1.0 metadata file is bundled with the FailureScope Croissant package on the umbrella record. Usage and field schema are documented in the included README and the paper appendix.

FailureScope 单轮语料库：从GSM8K、ARC-Challenge、MMLU、HumanEval、MBPP与IFEval中汇集的2664个任务，在参数规模覆盖1.3B至70B+的18个语言模型上完成评估。每条记录包含任务提示词、来源基准数据集、标准答案、各模型的通过率/失败率标签、分类学分配结果，以及留一模型聚类（leave-one-model-out, LOMO）簇ID。本数据集旨在支撑可复现的跨模型行为诊断。该数据集是FailureScope发布的三大组成部分之一（单轮、多轮、对抗式），相关标识可查阅总记录DOI：10.5281/zenodo.20037167。配套的Croissant 1.0元数据文件已随总记录中的FailureScope Croissant打包包一同提供。使用方法与字段架构已在附带的README文档及论文附录中予以说明。