Thunder-NUBench (Negation Understanding Benchmark) is a benchmark specifically designed to evaluate large language models’ (LLMs) sentence-level understanding of negation.
Unlike prior benchmarks that treat negation as a minor or syntactic feature, Thunder-NUBench introduces rich, manually curated sentence pairs and multiple-choice tasks that contrast standard negation with structurally similar distractors (e.g., local… See the full description on the dataset page: https://huggingface.co/datasets/thunder-research-group/SNU_Thunder-NUBench.
