LLM: RAFT Enhances RAG Robustness
In a collaborative effort between Microsoft and the University of California, Berkeley, researchers delve into the question: Can fine-tuning enhance the robustness of Retrieval-Augmented Generation (RAG) applications? Focused on evaluating the potential of small open Language Models (LLMs) such as Meta Llama 7B to match the capabilities of OpenAI's GPT-3.5, they introduce "Retrieval Augmented Fine Tuning (RAFT)" as a novel approach. RAFT revolves around training an LLM to filter out unhelpful documents retrieved during the process, aiming to refine the model's ability to generate accurate answers while incorporating chain-of-thought responses. Central to this technique is the inclusion of distractor documents alongside relevant ones, teaching the model to discern and disregard irrelevant information effectively. Key insights gleaned from the study highlight the efficacy of RAFT in enhancing the robustness of RAG applications: The incorporation of chain-of-thought (CoT) answers yields performance improvements of up to approximately 15%, underscoring the importance of contextual coherence in generating accurate responses. The addition of distractor documents contributes to the model's resilience, bolstering its ability to filter out noise and focus on relevant information. RAFT-trained Llama 2 7B demonstrates superior performance compared to GPT-3.5, particularly evident in specialized domains such as PubMed (Medical), showcasing its potential for domain-specific applications. The optimal number of documents included during training should align with the inference phase, ensuring consistency and efficiency in model performance. Overall, the findings underscore the promise of RAFT as a method to enhance the robustness and effectiveness of RAG applications, paving the way for improved performance and applicability across diverse domains. As researchers continue to explore and refine fine-tuning techniques, RAFT stands out as a valuable tool in advancing the capabilities of LLMs and unlocking new possibilities in natural language processing.