Core Concepts
Retrieval-Augmented Generation outperforms Fine-Tuning for developing G-LLM-based knowledge systems.
Abstract
The study compares Retrieval-Augmented Generation (RAG) and Fine-Tuning (FN) techniques for G-LLM models. RAG-based constructions show higher efficiency than FN models, with a 16% better ROUGE score, 15% better BLEU score, and 53% better cosine similarity. The study highlights challenges in connecting FN with RAG due to potential performance decreases. It emphasizes the advantages of RAG over FN in terms of hallucination and creativity. The research explores data preparation methods, model selection criteria, metric evaluation strategies, and training settings for both approaches.
Stats
Based on measurements shown on different datasets, we demonstrate that RAG-based constructions are more efficient than models produced with FN.
Outperforms the FN models by 16% in terms of the ROGUE score, 15% in the case of the BLEU score, and 53% based on the cosine similarity.
The average 8% better METEOR score of FN models indicates greater creativity compared to RAG.
Quotes
"Connecting FN models with RAG can cause a decrease in performance."
"RAG significantly improves hallucination compared to fine-tuned models."
"The best result was obtained using RAG Llama-2-7b base model with indexed datasets."