Money on the Mind: Do LLMs Get Selfish?

An NLP research project investigating whether monetary priming affects decision-making in Large Language Models (LLMs).

Research Question

This project explores whether exposure to money-related concepts influences LLM behavior in strategic decision-making scenarios, specifically testing if LLMs become more selfish or cooperative when primed with different framings of money.

Project Structure (will get bigger with time)

money_games/
├── NLP_group_20_Project_Runner.ipynb    # Reference notebook (end-to-end example)
├── requirements.txt                      # Project dependencies
└── src/
    ├── experiment/
    │   ├── main.py                      # CLI entry point to run experiments
    │   └── runner.py                    # Experiment orchestration logic
    ├── models/
    │   ├── config.py                    # Model constants
    │   ├── model_manager.py             # Model loading and management
    │   ├── open_source_model.py         # Open source model wrapper
    │   └── gemini.py                    # Gemini API wrapper
    ├── prompts/
    │   ├── prompt_builder.py            # Prompt construction utilities
    │   └── configs/                     # Constants for the prompts
    │       ├── money.py                 
    │       └── games.py                
    └── analysis/
        └── token_probs.py               # Token probability analysis tools

Setup (with virtual environment)

It's recommended to use a virtual environment.

Windows (PowerShell):

python -m venv venv
./venv/Scripts/Activate.ps1
pip install -r requirements.txt

macOS/Linux (bash/zsh):

python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Quickstart (CLI)

Run using module execution with the venv's Python:

• Run both open-source and Gemini experiments (saves results to results/experiment_results.csv):

python -m src.experiment.main --hf-token YOUR_HF_TOKEN --gemini-key YOUR_GEMINI_KEY

• Run only open-source models:

python -m src.experiment.main --experiment open_source --hf-token YOUR_HF_TOKEN --output open_source_results.csv

• Run only Gemini:

python -m src.experiment.main --experiment gemini --gemini-key YOUR_GEMINI_KEY --output gemini_results.csv

• Dry run (print plan without executing):

python -m src.experiment.main --experiment both --hf-token YOUR_HF_TOKEN --gemini-key YOUR_GEMINI_KEY --dry-run

Your example

python -m src.experiment.main --experiment open_source --hf-token YOUR_HF_TOKEN --gemini-key aa --output open_source_results.csv

Arguments

• --experiment|-e: open_source | gemini | both (default: both)
• --hf-token|-t: HuggingFace token (required for open_source or both)
• --gemini-key|-g: Gemini API key (required for gemini or both)
• --output|-o: Output CSV filename (default: experiment_results.csv)
• --output-dir: Output directory (default: results)
• --models: Optional list to override models (uses defaults in src/models/config.py if omitted)
• --clear-results: Clear in-memory results before running
• --dry-run: Show configuration without executing
• --verbose|-v: Verbose error output

Environment variables are set automatically for the run, but you can also pre-set them:

• HUGGINGFACE_TOKEN
• GEMINI_API_KEY

What the CLI prints/saves

• A tabular summary of results: model, prefix_type, paraphrase_index, response
• Counts per model/prefix and overall totals
• Decision token probability analysis for open-source models (if available)
• Response pattern analysis (samples by prefix and by model)
• Results saved as CSV to --output-dir/--output

Notebook (reference)

You can still follow the full end-to-end example in NLP_group_20_Project_Runner.ipynb. The CLI mirrors the same flow (setup → run → summarize → analyze → save) and is better suited for automation.