Synthetic Dialogue Dataset for Romance Scam Detection

This is a synthetically generated conversation dataset designed for research on romance scam detection in online dating platforms. It contains 1,000 multi-turn conversations — 500 simulating romance fraud and 500 simulating genuine dating interactions — each consisting of 20 alternating messages between two LLM-driven personas. Purpose: To provide labelled training data for building protective AI systems that can identify romance scam patterns in real-time messaging. The dataset addresses the scarcity of publicly available romance scam conversation data, which is difficult to collect ethically from real victims. Nature: All conversations are generated locally using Meta LLaMA 3 (8B) via Ollama, with no external API calls or real user data involved. Scam conversations follow a documented three-phase romance fraud arc-trust-building, love-bombing, then financial exploitation — while legitimate conversations cover everyday dating topics (hobbies, travel, food, pets). Each persona maintains perspective-aware conversation history to produce coherent, contextually grounded dialogue. Scope: The dataset covers a single romance scam archetype (overseas worker/military deployment leading to financial requests) in English, with four fixed character personas across a uniform 20-message conversation length. It is scoped for binary classification (scam vs. legitimate), temporal pattern analysis, reinforcement learning reward modelling, and federated learning experiments within the broader project. It is not intended as a comprehensive representation of all online fraud types, languages, or conversation dynamics found in real-world dating platforms.