matthewwall369/dense

Dense is a research tool for people who build and study language models, the kind of AI systems behind chatbots. The project describes itself as being like a code editor, but built specifically for training, testing, comparing, and visualizing these models, so a researcher does not need to write custom setup code for every experiment. The tool has a dashboard that shows live hardware usage like GPU and memory, along with training metrics such as loss and learning rate as an experiment runs. It comes with several ready made datasets for testing, such as TinyShakespeare and WikiText, and it supports different ways of breaking text into tokens for a model to read. Users can plug in different model architectures, including well known ones like Transformer as well as newer alternatives, and the tool will pick them up automatically without extra setup. Once a model is training, you can pause and resume it, watch live metrics update over a network connection, and save checkpoints along the way. There is a playground area for generating text from a trained model with adjustable settings, and an analysis section for inspecting the internal structure of a model and comparing different architectures against each other on a leaderboard. Results can be exported in several formats including spreadsheets, web pages, and ready made notebooks. Getting started requires Python and Node.js installed locally, with an optional graphics card for faster training. The backend is written in Python using FastAPI and PyTorch, while the frontend is a React and TypeScript interface. The project also ships with Docker support so it can be run in one command without installing anything locally. It is released under the MIT license and is described by its author as a portfolio project demonstrating a full research tool built end to end.