Language models such as GPT-3 have demonstrated impressive general knowledge and understanding. However, they have limitations when required to handle specialized, niche topics. Therefore, a deeper domain knowledge is necessary for effectively researching specific subject matter. This can be equated to asking a straight-A high school student about quantum physics. They might be smart, but not equipped with specialized knowledge in that area.
Addressing this issue, a group of researchers at UC Berkeley have put forward an innovative approach, referred to as RAFT (Retrieval Augmented Fine Tuning). It is intended to serve as a conduit between generalized artificial intelligence (AI) and highly specific expertise. Essentially, RAFT would arm generalist language models with specialized knowledge and documentation.
In traditional methods, tools like GPT-3, while impressive in their breadth of abilities, falter when it comes to domain-specific knowledge. This is where RAFT comes in through a training process that simulates an open-book exam:
1) It trains on question-answer pairs from the specialized domain.
2) It is presented with a mix of relevant “oracle” documents and irrelevant “distractor” documents.
3) The model then learns to sift through these data, citing relevant quotes and building multi-step reasoning.
This process teaches the model to focus its attention and understanding on subject-specific content. It was evaluated on coding, biomedicine, and general question-answering benchmarks, where it showed substantial improvements over traditional fine-tuning methods.
The results confirmed RAFT’s superiority over existing models across a range of specialized domains. When tested on datasets such as biomedicine literature from PubMed, general questions from HotpotQA, and coding benchmarks from HuggingFace and TorchHub, RAFT consistently outperformed other models. The most notable improvement was a 35.25% increase on the HotpotQA and a 76.35% jump on TorchHub’s coding evaluation. When compared with GPT-3.5, RAFT demonstrated a clear advantage in leveraging the provided context and specialized knowledge to solve complex questions accurately.
RAFT’s arrival signifies a significant change in unlocking domain-specific knowledge for AI language models. This could give rise to digital assistants and chatbots that can provide expert guidance in various domains, from genetics to gourmet cooking. By integrating general reasoning with specialized subject matter competency, RAFT has the potential to transform language AI from ‘jacks of all trades’ to experts on specific subject matters, thus opening up new possibilities in industries like healthcare, law, science, and software development. RAFT stands out as an essential development in equipping AI with the ability to keep pace with or potentially overtake human expertise in every conceivable knowledge domain.
