Digital Lexicon Hub

Further training of a pretrained model on a smaller, task-specific dataset to change its behavior, style or accuracy for a particular use.

Fine-tuning can make a smaller, cheaper model perform a narrow task reliably and shorten prompts, but it needs quality data, evaluation and maintenance, so it should be weighed against prompt engineering and RAG first.

A claims team collects 2,000 reviewed emails labeled into 12 claim types, runs evals on the base model to set a baseline, fine-tunes with LoRA and adopts the tuned model only if it beats the baseline on a held-out set of 300 emails.

Fine-tuning teaches patterns, not reliable new facts, so frequently changing knowledge is better served by retrieval. Small or biased datasets cause overfitting or loss of general skills, and each base model upgrade may require retraining.

Full fine-tuning updates every weight and can need many times more GPU memory than the model weights alone. Parameter-efficient methods such as LoRA freeze the original weights and train small low-rank update matrices that can be swapped per task. Supervised fine-tuning uses pairs of prompts and ideal responses, while preference methods such as DPO learn from ranked outputs. Retrieval-Augmented Generation (RAG) supplies facts at query time, whereas fine-tuning changes how a Large Language Model (LLM) behaves.