LLM Adaptation: Fine-tuning as a strategy for medium scale projects

This white paper analyzes two approaches for customizing Large Language Models (LLMs) as alternatives to Retrieval-Augmented Generation (RAG): full fine-tuning of small open-source LLMs and Low-Rank Adaptation (LoRA) of medium-sized LLMs.

The research examines technical requirements, resource implications, and economic feasibility through practical experimentation. Using infrastructure ranging from local servers to cloud-based solutions, the study tested models like FLAN T5 and Mistral 7B.

While small model fine-tuning proved efficient but limited in capability, LoRA adaptation of medium-sized models showed promise as a balanced approach for organizations with constrained resources seeking LLM customization.

Download the white paper