ojayy/logical-sata

LOGICAL-SATA LOGICAL-SATA is a reading-comprehension benchmark for compound answer reasoning. Each instance has a paragraph, a question, and four candidate answer options. Every option joins two atomic answers under an explicit logical operator, AND, OR, or NEITHER/NOR. Exactly one option is valid per instance. The dataset is built to isolate logical composition from comprehension, so a model can get every atomic judgment right and still fail if it can't combine them correctly… See the full description on the dataset page: https://huggingface.co/datasets/ojayy/logical-sata.