This app is a part of the project deliverables for the BlueDot Impact course on AI Safety Fundamentals.
Read the experiment report: AI Debate for Defeasible Reasoning.
- Model Setup
- Scenario
- Debate Setup
- Baseline
- Debate
- Judge
How the Debate Works:
- First, review the scenario details and question from the BoardgameQA dataset, or create your own
- Assign positions to Debater A and B from the available options
- Start the debate by having models argue their assigned positions
- A judge model will evaluate the debate and determine the winner
You can change the scenario, reset the debate, and download the result at any time. The prompts are also editable to customize the debate experience. Make sure to keep the variables enclosed in curly braces ({}). The default prompts instruct the models to "think internally" which you can see in the Thinking Process section on each generated responses. Enjoy!