RAFT v3.1 open-sources retrieval-augmented fine-tuning to clone human personas

jessi_cata · x · 2026-09-15

RAFT (Retrieval-Augmented Fine-Tuning) v3.1 is out: an open-source pipeline that trains persona models on your past writings and dated interaction logs. Relevant corpus excerpts dated before each exchange are retrieved as memories and used as training targets, teaching the model how the author drew on past work in conversation. Supports X, Substack, LessWrong and more as sources, offers a CLI/wizard workflow from corpus to serving, and runs on OpenAI or your own GPU — a demo persona trained on gwern's writings is included.

Original post →

More from coding & agent

coding & agent channel →