Improving RAG effectiveness with Retrieval-Augmented Dual Instruction Tuning (RA-DIT)

Meta's AI Research team has developed a method called RA-DIT (Retrieval-Augmented Dual Instruction Tuning) to enhance Large Language Models (LLMs) by incorporating retrieval features without the costly modifications typically required during pre-training or post-training integration. RA-DIT employs a dual fine-tuning process that separately optimizes the language model and the retriever, improving the model's ability to utilize retrieved information and refine the retriever's search capabilities. This approach involves creating a fine-tuning dataset with Q/A pairs and utilizing the LM-Supervised Retrieval (LSR) method, enabling the language model to produce accurate predictions even with incorrect data retrievals. RA-DIT was tested on knowledge-intensive and commonsense reasoning tasks, demonstrating superior performance compared to both the untuned LLAMA 65B model and the LLAMA 65B REPLUG, as well as showing improvements over base LLAMA 65B models on multiple evaluation datasets. The RA-DIT approach enhances the LLM's accuracy and context-awareness by fostering a more efficient data retrieval process, ensuring that generated answers are both relevant and informed.