agentlans/synthetic-social-dialogues

Synthetic Social Dialogues

Synthetic Social Dialogues is a collection of over 8 000 synthetically generated, multi-turn dialogue scenarios designed to train and evaluate Large Language Models (LLMs) on nuanced conversational dynamics, role-playing, and adherence to complex social constraints.

Each sample models a realistic interaction between two speakers under controlled psychological, relational, situational, and medium-based variables.

Dataset Summary

• Size: 8 000+ rows
• Format: JSON-structured fields containing metadata, execution prompts, and generated dialogues.
• Language: English
• Generation Model: google/gemma-4-E4B-it

Dataset Structure

Each row in the dataset contains three primary fields:

• context: A JSON string containing randomly generated structural metadata for the scenario. This includes character profiles (generated via the Faker library) tracking attributes like name, role, current mood, stress level, energy, and dominant traits, as well as situational constraints (intent, relationship type, medium, cultural norms, stake level, and knowledge asymmetry).
• prompt: The full, structured system instructions fed to the LLM to generate the conversation based on the variables in the context.
• output: A raw JSON string containing a short description of the inferred scenario and the final multi-turn dialogue breakdown.

Feature Schema

{
  "context": "STORY_METADATA_JSON",
  "prompt": "LLM_SYSTEM_PROMPT_STRING",
  "output": {
    "inferred_scenario": "SCENARIO_DESCRIPTION_STRING",
    "dialogue": [
      { "speaker": "SPEAKER_NAME", "line": "DIALOGUE_LINE_STRING" }
    ]
  }
}

Creation Methodology

1. Scenario Categorization: Scenarios are randomly selected from one of three high-level tracks: corporate_professional, personal_crisis_or_intimate, or casual_social.
2. Variable Assignment: Relationships, settings, mediums (e.g., in-person, email, Slack), communication goals, and cultural baselines are procedurally mapped to the scenario.
3. Persona Generation: Profiles including mood, stress, energy, and traits are assigned. Personal identities (names and genders) are generated using the Faker Python package.
4. Inference: The bundled parameters are structured into a strict prompt directive requiring natural, non-cliché conversational outputs matching the chosen medium rules and outputting in raw JSON format.

Limitations

• Length & Detail: Some generated dialogues can be relatively brief or occasionally fail to fully integrate every nuanced sub-attribute specified in the prompt.
• Verbal Language: Generated dialogues often lack non-verbal language, actions, and gestures that are vital in real-world communication.
• Alignment Bias: There is a noticeable absence of explicit profanity, hostile language, or intense verbal friction. Toxic behaviours like malicious gaslighting or verbal abuse resolve far more cleanly and politely than real-world human conflicts.
• Cultural Homogeneity: Despite the presence of a cultural_norms steering variable, scenarios strongly favour Western and American cultural perspectives and corporate structures. While specialized modern niches (like youth and gaming subculture) are present, other global cultural frameworks are underrepresented.
• Static Psychological States: Real-world human conversational emotional states are dynamic. In this dataset, character moods and psychological states remain static from the beginning to the end of the dialogue block.

Licensing

This dataset is licensed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. You are free to share, copy, modify, and adapt the material for any purpose, including commercially, provided you give appropriate credit.