composable-models/llm_multiagent_debate ? reverse-engineered prompt

Build me a Python project that recreates this multiagent debate setup for language model reasoning and factuality. I want scripts that can run debate based answer generation for a few benchmark style tasks, including math problems, grade school math, biographies, and MMLU style questions, plus separate evaluation scripts for the ones that need scoring. Keep it simple and research oriented, so I can run each task from its own folder and generate outputs with one command. Please make the code easy to follow, include clear README instructions for downloading the datasets and running the experiments, and structure it so I can plug in model calls and compare debate results against direct answers. If anything depends on current package versions or dataset details, look up the latest docs online if needed.