Circular Bias Detection (CBD) Dataset and Evaluation Protocols (v3 / v3.1)

We release a protocol-aware dataset for circular bias detection across time periods, evaluation protocols, and resource constraints. The dataset includes Sleuth-formatted CSVs (v3), augmented variants (v3.1), period-level MTL training JSONL with heuristic labels (PSI/CCS/RhoPC/Bias), and scripts for reproducible augmentation, labeling, and inference. It supports both discriminative multi-task learning and dialog-based explanatory models, forming a detection→explanation→remediation pipeline.