microsoft/tinytroupe

TinyTroupe is a Python research library from Microsoft that lets you simulate groups of people with distinct personalities, backgrounds, and goals. Each simulated person (called a TinyPerson) can hold conversations, respond to situations, and act within simple virtual environments. The underlying intelligence comes from large language models, specifically GPT-4 and its successors, which generate the behavior for each simulated character. The purpose is to help researchers and product teams understand how different kinds of people might react to something, without needing to recruit actual participants. You define personas (for example, a 45-year-old physician, a college student, a small business owner), place them in a simulated scenario, and observe what they say and do. The library also includes tools to statistically compare simulation results against real-world survey data, to check whether the simulations actually reflect genuine human behavior. Practical uses described in the documentation include testing digital advertisements with a simulated audience before spending money on them, generating realistic test inputs for chatbots or search engines, producing synthetic training data, and running virtual focus groups to get product feedback from specific persona types. You can also read project proposals to the simulated group and receive feedback from the perspective of particular professional roles. The library is experimental and under active development. The API changes frequently as the project matures. It requires an OpenAI API key to run, and there is limited experimental support for locally hosted models via Ollama. The project ships with Jupyter Notebook examples covering a range of scenarios. Microsoft notes that TinyTroupe is intended for research and simulation only, and users are responsible for how they use the generated outputs. The full README is longer than what was shown.