Translation: Recently, researchers at Meta have proposed a lightweight fine-tuning method called RA-DIT to enhance the knowledge retrieval capabilities of language models. This method involves a two-stage tuning process: the first stage improves the language model's ability to utilize retrieved information, and the second stage optimizes the retriever to provide more relevant content. Experimental results indicate that RA-DIT65B outperforms existing models in knowledge-intensive zero-shot and few-shot tests. It also significantly improves performance on tasks that require high levels of knowledge utilization and contextual understanding. The study demonstrates the effectiveness of RA-DIT's lightweight tuning for retrieval-enhanced language models, particularly in scenarios that require access to large-scale knowledge sources.