Skip to main content

Synthetic Dialogue Dataset for Romance Scam Detection

Ranaweera, Kanishka ; Ghari Neiat, Azadeh (2026) — Harvard Dataverse

View Original Source (opens in a new tab)

DOI: 10.7910/dvn/o317px (opens in a new tab)

Type:   Dataset
Tags:   Dating App Conversations, Fraud Classification, Large Language Models, Natural Language Processing, Romance Scam Detection, Social Engineering, Synthetic Dialogue
Citation tools
Download RIS Download BibTeX 


              
              

            

          

        


      

      

    

        
Transparency

        
Evidence and review status

        
This page contains AI-generated content. No human content review or subject-matter-expert review is recorded.

    


    

                    

                
Source basis

                

                    Downloaded PDF
                

            

                    

                
Source updates

                

                    Not checked
                

            

                    

                
AI-generated page content

                

                    Yes
                

            

                    

                
Automated checks

                

                    Passed
                

            

                    

                
Administrative approval

                

                    Yes
                

            

                    

                
Human content review

                

                    Not recorded
                

            

                    

                
Subject-matter-expert review

                

                    Not recorded
                

            

            


    

        
            How this was prepared
        
        

                            

                    
Source basis

                    

                    
RSRC downloaded and privately stored a copy of the paper for internal analysis. The PDF is not offered to viewers from this page.


                                            
    
                                                            
  • 
                                    PDF added to RSRC:
                                    
                                
    • 
                                                    

                    
                                        

                

                            

                    
Source updates

                    

                    
No completed source-status check for the current DOI is recorded.


                    
                                        

                

                            

                    
AI-generated page content

                    

                    
AI-generated research notes displayed on this page: Synopsis, Identified gaps, Methods, Limitations, Future work. The paper itself is not described as AI-generated.


                                            
    
                                                            
  • 
                                    Document analysis recorded:
                                    
                                
    • 
                                                            
  • 
                                    Synopsis generation recorded:
                                    
                                
    • 
                                                            
  • 
                                    Page record updated:
                                    
                                
    • 
                                                    

                    
                                        

                

                            

                    
Automated checks

                    

                    
The current, source-bound synopsis passed the recorded versioned publication checks.


                                            
    
                                                            
  • 
                                    Checks completed:
                                    
                                
    • 
                                                    

                    
                                            

                            
                                View passed checks (3)
                            
                            
    
                                                                    
  • Length, completeness, repetition, refusal, boilerplate, and active-markup screening
    • 
                                                                    
  • Numerical claims checked against the available source text
    • 
                                                                    
  • English-source lexical grounding check
    • 
                                                            

                        

                                        

                

                            

                    
Administrative approval

                    

                    
An authenticated administrator approved the bibliographic record for public Library display. This is not a review of every research claim.


                                            
    
                                                            
  • 
                                    Approved for public display:
                                    
                                
    • 
                                                    

                    
                                        

                

                            

                    
Human content review

                    

                    
No human review is recorded for the AI-generated content displayed on this page.


                    
                                        

                

                            

                    
Subject-matter-expert review

                    

                    
RSRC has not recorded review of this content by a subject-matter or methods expert.


                    
                                        

                

                    

    


    

        Review-state definitions

        Found a possible error?
        Request a correction.
    



    



              

        
                  

            
Synopsis

            
The publication presents a synthetic dialogue dataset designed to support romance scam detection research. It describes 1,000 conversations generated with two paired personas per scenario: scam (Marcus and Sarah) and legitimate (Jake and Emily). Each conversation comprises 20 messages, resulting in 20,000 total messages, stored as JSONL objects. The scam conversations follow a three-phase arc—warm early exchanges, mid-stage love-bombing and hardship claims, and late-stage financial requests—while legitimate conversations focus on ordinary topics and reciprocal curiosity. The data are produced by a local LLM setup using Ollama, with system prompts guiding each persona’s behavior, and the conversations are archived with a labeled flag indicating scam or legitimate origin. The dataset details, generation workflow, and intended uses are documented extensively. Each line in the dataset includes a conversation_id, label (1 for scam, 0 for legitimate, with 500 of each), a messages array containing sender, message text, and turn number, and a source tag (synthetic_scam or synthetic_legit). The authors outline potential applications such as training binary classifiers, reinforcing RL environments for real-time intervention, federated learning experiments, NLP feature testing, and temporal pattern analysis. They also enumerate limitations, including the synthetic origin, a fixed four-character set, a single scam archetype, uniform 20-message length, and English-only content, noting that real-world patterns may differ. The work emphasizes ethical use for defensive purposes and cautions that models trained only on synthetic data may not generalize to all real-world variants.

          

        
                  

            
Identified Gaps

            
The dataset does not represent authentic user conversations, diverse scam playbooks, multilingual interactions, variable scam timelines, or platform metadata. Its fixed four-persona design and balanced classes leave uncertainty about how well models trained on it generalize to real-world romance-scam detection.

          

        
                  

            
Methods

            
The dataset uses a Python script and Ollama-hosted Meta LLaMA 3 8B to generate 1,000 English dating-app dialogues. Two prompted personas alternate for 20 turns per conversation: 500 scripted scam dialogues and 500 legitimate dialogues. Messages are post-processed to remove quotation marks and character-name prefixes, then stored as labeled JSONL records.

          

        
                  

            
Limitations

            
All dialogues are synthetic and may differ from authentic interactions. Only four fixed personas, one romance-scam archetype, English text, and exactly 20 messages are represented. The dataset lacks timestamps, profiles, multimedia, and other platform metadata; its balanced class distribution is unrealistic; and dialogue quality depends on LLaMA 3 8B and possible safety refusals.

          

        
                  

            
Future Work

            
Supplement synthetic data with ethically obtained real-world conversations for production classifiers. Extend coverage to additional scam variants, communication styles, cultural backgrounds, languages, variable conversation lengths, and non-textual platform metadata.

          

              

    
    

        
Related Research Insights

        

            See how this publication connects to RSRC's living evidence syntheses through current citations and research-topic mapping.