FedHeartShield Reviewed Romance Scam Corpus

A synthetic, human-reviewed corpus of romance-scam and genuine conversations for scam-detection and online-safety research. The full corpus holds 404 conversations, balanced 202 scam and 202 genuine, with matched and counterfactual benign controls and turn-level risk, stage, and manipulation annotations. It is released in two tiers: this open tier provides the datasheet, data dictionary, validation summaries, and 10 bounded illustrative examples under CC BY-NC 4.0, and the full dialogues are available on request (see ACCESS.md). All data is synthetic, with no authentic victim data and no personally identifying information.

Known limitation: the scam and genuine classes are separable by a writing-style (register) fingerprint left by the generation process rather than by scam-relevant evidence (held-out AUC 0.84–1.0; see SEPARABILITY.md). The corpus suits research on corpus-validity auditing and de-shortcutting methods, and capability or calibration work, rather than measuring detection effectiveness. Released for non-commercial research under CC BY-NC 4.0.