Fine-tuning

Overview

Fine-tuning continues training a pre-trained foundation model (GPT, Llama, Qwen, etc.) on a domain-specific or task-specific dataset to improve accuracy or output consistency. Full fine-tuning updates all parameters; parameter-efficient methods like LoRA and QLoRA update only a small adapter, making the process feasible on consumer hardware.

Fine-tuning vs RAG

Fine-tuning excels at internalizing writing style, tone, and domain vocabulary. RAG is better for retrieving up-to-date factual knowledge. Production systems often combine both: RAG for retrieval and fine-tuning for style and domain alignment.