<p>Modern artificial intelligence (AI) models, such as GPT and others, already possess extensive knowledge due to training on large text corpora. However, this knowledge is limited to the data on which the model was trained and does not update over time. To address this issue, two methods are used: finetuning and Retrieval Augmented Generation (RAG).</p>
<p>Finetuning is additional training of the model on a specialized dataset. This method enhances AI knowledge in a specific area but is not always effective for general knowledge expansion.</p>
<p>RAG approaches the task of updating knowledge differently. It works by adding relevant information found in a vast database to the model's query. This allows the model to generate responses that are better grounded in facts and current information.</p>
<p>Recent studies show that RAG is more effective than finetuning in the task of embedding new knowledge into AI. While finetuning does improve model performance compared to the baseline, RAG has a significant advantage.</p>
<p>📝 Paper: <a href="https://arxiv.org/abs/2312.05934">https://arxiv.org/abs/2312.05934</a></p>;
<p>#ai #gpt #llm #rag_vs_finetuning</p>
<p>МР</p>
