llm-colosseum

Why we included this project

LLM evaluation usually means ticking off multiple-choice questions, but this project stages the test as a live Street Fighter III match. Two models fight each other in real time, and the winner is the one that reads the changing screen and reacts to its opponent fast enough to matter. The repo ships a text robot that feeds the model a written description of the game and a vision robot that hands it a screenshot, so you can see how text-only and multimodal models behave in the same spot. It also runs local inference through Ollama, which lets small teams run their own head-to-head matchups without paying for API calls on every experiment.