Raft

RAFT, short for retrieval augmented fine tuning, is an open source tool that trains a persona model from a person's past writing and dated interaction logs. The result is a model that imitates how an author writes and draws on earlier work when answering. It runs as a local project with the user's choice of model, driven by an interactive terminal wizard or single commands. The method has three parts. Past writing is chunked, dated and embedded in a vector store. For each training conversation, only writing dated before that exchange is retrieved, and a reasoning trace connects the recalled memories to the real reply. At inference, relevant memories are retrieved and prefilled into the model's thinking before it answers. Importers cover LessWrong and the EA Forum, X and Bluesky, Substack, blogs and feeds, web pages, PDFs, local files and chat logs, each available as a command such as raft fetch. Sources can be assigned a role as conversations, grounding documents, or both. Training can run through OpenAI or on a GPU through opbdh, and the workflow includes benchmarks and test questions for evaluating the persona, then a chat with the trained model. A demo clone called gwernbot is built from fifteen years of a public forum commenter's writing and needs no API keys. The site states no consent requirement or usage restriction for cloning other people's writing, and publishes no pricing. Documentation is at version 3.1.0.