allentdan/fastchat

FastChat is a complete toolkit for working with AI chatbots, it lets you train them, run them, and measure how well they work. The most visible thing it powers is Chatbot Arena, a website where you can chat with 50+ different AI models and vote on which responses you like better. Those votes feed into a public leaderboard that ranks which models are best. At its core, FastChat does three main things. First, it includes pre-trained chatbot models you can download and use immediately, most famously Vicuna, an AI assistant that was fine-tuned to work like ChatGPT. Second, it provides an easy way to run these models yourself, either through a simple text chat in your terminal or through a web interface that feels like talking to ChatGPT. Third, it offers tools to evaluate how good your chatbots are by having them answer challenging questions and having humans (or other AI models) judge the quality of their responses. The platform is flexible about where and how you run things. You can run a model on your laptop's CPU, on a gaming GPU, on multiple GPUs at once, or even on specialized hardware like Apple's chips or Intel accelerators, it adapts to what you have. It also provides an API that mimics OpenAI's interface, so if you've written code that talks to ChatGPT, you can swap in FastChat with minimal changes. This makes it practical for people building applications who want to self-host their AI instead of paying OpenAI's fees. Researchers and companies use FastChat to train custom chatbots on their own data, then serve them to users. Because it's open-source, anyone can see how the best models were built, and the included evaluation tools let you objectively compare different approaches. The combination of real conversations from Chatbot Arena (over 100,000 human votes collected) plus the ability to fine-tune models makes it a complete platform for the AI chatbot development cycle.