• Sample_Persona_Meta/: Code for generating meta personas from US Census Bureau data

  • get_dist/: Probability sampling functions for demographic information (age, sex, race, education, income, etc.)
  • collect_data/: Data collection notebooks from US Census Bureau
  • sample_marginal.py: Main script for generating meta personas

• Gen_Sim/: Core simulation framework

  • Persona_Generation.py: Generates detailed personas from meta personas
  • Opinion_Simulation.py: Runs opinion simulations with generated personas
  • Persona_Calibration.py: Calibrating persona generation to avoid sampling bias.
  • configs/: Configuration files for different models and simulation parameters
  • model_configs/: Model-specific configurations for various LLMs
  • run_persona_gen.py: Script for generating personas
  • run_simluation_topic.py: Script for running topic-based simulations
  • run_simluation_slurm.py: SLURM cluster execution script

• Prompts/: Prompt templates for persona generation and opinion simulation

  • Persona_Generation/: Prompts for creating detailed personas. We also include prompts we used for experimentation. For the simulation experiments in the paper, we used the following persona types:
    • objective_table_persona/: Structured objective persona with demographic tables
    • subjective_table_persona/: Structured subjective persona with opinion tables
    • descriptive_persona/: Natural language descriptive personas
  • Opinion_Simulation/: Prompts for opinion simulation tasks
    • fast_user_instruction.json: Fast simulation without chain-of-thought
    • multi_turn_*.json: Multi-turn conversation simulation prompts
    • reason_user_instruction.json: CoT-based opinion generation

• Simulation_Result/: Output directory for simulation results

  • Contains results organized by topic and persona type

First, generate meta personas from US Census data:

cd Sample_Persona_Meta
python sample_marginal.py

This creates demographic samples for all US states based on Census Bureau data.

Generate detailed personas using LLMs:

cd Gen_Sim
python run_persona_gen.py

This creates the three persona types (objective table, subjective table, descriptive) for each state.

First, create simulation questions in the Simulation_Result/ directory, register the questions as a file called context.json. See the context.json file in the Simulation_Result/ directory for an example.

Run opinion simulations on various topics:

cd Gen_Sim
python run_simluation_topic.py

This simulates opinions across different topics using the generated personas.

• configs/persona_config.json: Configuration for persona generation
• configs/simulation_config.json: Configuration for opinion simulations
• model_configs/: Model-specific configurations for different LLMs

• Llama-3.1-8B-Instruct
• Llama-3.1-70B-Instruct
• Athene-70B
• Mixtral-8x7B-Instruct-v0.1
• Nemotron-70B-Instruct
• Qwen2.5-72B-Instruct

You can register new models in the model_configs/ directory.