Master Debate

A Python application that orchestrates debates between multiple LLMs, each embodying a different philosopher's personality.

Features

• 2-5 Philosophers debate any given issue
• 20 Famous Philosophers to choose from (Socrates, Plato, Aristotle, Kant, Nietzsche, etc.)
• LLM Integration with Ollama or compatible APIs
• Automatic Model Discovery from /api/tags, plus a Custom option for any name
• Streaming Output — responses appear as they are generated, with a working Stop button
• Consensus Detection by an independent judge, checked every 2 turns and always on the final turn
• Independent Judge Model for consensus checks and summaries, run at temperature 0 so the same debate gets the same verdict
• Time & Turn Limits — the debate stops at the configured duration or after 20 turns
• Clear Failure Reporting — an unreachable backend or a missing model is shown, not swallowed
• Crash-Safe Storage — settings and saved debates are written atomically, so a reader never sees a half-written file
• Streamlit Web UI for easy interaction

Setup

1. Install Dependencies

pip install -r requirements.txt

2. Setup LLM Backend

Ensure you have an LLM API server running. Options:

Ollama (Recommended)

# Install Ollama from https://ollama.com
ollama pull llama3
ollama pull phi4

Compatible API Endpoints

The app expects endpoints at /api/generate and /api/chat (Ollama-compatible format).

3. Run the App

streamlit run app.py

How It Works

1. Select Philosophers: Choose 2-5 philosophers from the dropdown
2. Enter Issue: Type the philosophical issue to debate
3. Set Time Limit: Configure how long the debate may run (default 30 minutes)
4. Choose Models: Pick from the models discovered on your server, or type a custom name
5. Choose a Judge: Optionally pick a separate model for the consensus/summary calls
6. Start Debate: Watch the arguments stream in turn by turn; press Stop to end early
7. Consensus Check: Every 2 turns — and always on the final turn — the judge model evaluates whether consensus was reached
8. Review Results: See the outcome, statistics, summary and full history

Philosophy Personalities

Each philosopher has a unique personality defined by their core philosophical principles:

• Socrates: Dialectical questioning, "knowing that I know nothing"
• Plato: Theory of Forms, analogies and allegories
• Aristotle: Empirical observation, logical causality
• Descartes: Radical doubt, "I think therefore I am"
• Nietzsche: Will to Power, critique of morality
• Rawls: Veil of ignorance, fairness as justice
• And 10+ more!

Configuration

• LLM API URL: Default http://localhost:11434
• Model Names: Discovered from the provider's /api/tags; if the server is unreachable the built-in list is used, and Custom… accepts any name
• Judge Model: Model used for the consensus and summary calls (defaults to the selected LLM model)
• Judge Temperature: 0.0 — the judge only classifies and summarises, so it runs greedily
• Per-philosopher overrides: Off by default; enables a provider/model per philosopher
• Time Limit: The debate stops once this duration elapses (default 30 minutes)
• Max Turns: 20 turns maximum per debate
• Consensus Check Interval: Every 2 turns, plus the final turn (judging is a full extra model call)
• Temperature: 0.7 (balanced creativity/rationality) for the debating philosophers
• Response Length: 512 tokens per turn
• Request Timeouts: 5 seconds to connect, 10 minutes to read a response
• Python: 3.11 or newer

Storage

• debate_settings.json holds the saved configuration and debates_data/ holds finished debates. Both paths can be redirected with the MASTER_DEBATE_SETTINGS_FILE and MASTER_DEBATE_DEBATES_DIR environment variables.
• Writes go to a temporary file and are moved into place with os.replace, and settings access is serialised by a lock. Every Streamlit session in a process shares these files, so a file is never left half-written, but the last session to change a setting still wins.

Tips

• Model Quality: Better philosophical reasoning with larger models (Llama3 70B, Phi-4, etc.)
• Consensus: Tough philosophical issues rarely reach consensus - that's the point!
• Debate Quality: Try contrasting philosophies (e.g., Rawls vs. Nozick on justice)

Testing

pip install -r requirements-dev.txt
pytest
ruff check .

The suite covers the debate engine, the streaming client, the background debate runner, the philosopher roster, and the Streamlit app (via streamlit.testing.v1.AppTest), so it runs without a network connection or an LLM backend.

License

MIT License — see LICENSE for details. Feel free to modify and extend!