RAFT (Retrieval Augmented FineTuning) presents a novel training approach to incorporate new knowledge into Large Language Models for domain-specific tasks.

• RAFT enables LLMs to discern and dismiss distractor documents during open-book questioning.
• It sharpens the models’ reasoning abilities with chain-of-thought-styled responses.
• Showcases impressive performance improvements across several datasets.
• RAFT’s open-source code is available to foster further innovation in the AI community.

The methodology’s focus on contextual question answering and reasoning extends the capabilities of LLMs significantly. RAFT’s deployment could pave the way for advanced AI services tailored to specialized domains, including healthcare and legal applications.