t

Each problem presents a set of variable equality/inequality premises and asks the model to identify which conclusion necessarily follows. Unlike knowledge-based benchmarks, LogicMark contains no facts a model could have memorised from pretraining.

Every problem is generated fresh from abstract variable names (a, b, c, ...), so a model cannot pattern-match to training data - it must… See the full description on the dataset page: huggingface.co/datasets/AxiomicLabs/LogicMark.