PhyseaWiki How AI actually works physea.ai →

AI stacks

What does a fine-tuning stack involve, and when should you use it?

The fine-tuning stack is a base model, a labeled dataset, a training method (LoRA for cheap, full fine-tuning for thorough), and an eval that proves the result improved. Fine-tune to change a model's behavior or style, not to teach it new facts. For facts, RAG is usually the better tool.

Last updated 2026-07-25 · Physea Labs

Fine-tuning is teaching an existing model new behavior by training it further on examples. The stack has four parts: a base model to start from, a dataset of examples that show the behavior you want, a training method that does the updating, and an eval that proves the result is actually better. Skip the eval and you have no way to tell whether you improved the model or quietly broke it.

Base model the starting point Dataset examples of the behavior Training method LoRA or full fine-tune Eval proves it improved
Skip the last box and the first three have no way to prove they actually improved anything.

In practice this is not a one-way pipeline. A bad eval sends you back to the dataset (more examples, better-labeled ones) or the method (a different LoRA rank, more training steps), rather than to the base model — which is usually the one part of the four you do not touch.

The method is where the real choice lives. Full fine-tuning updates every weight in the model; it is the most thorough option and the most expensive in compute and memory. LoRA, low-rank adaptation, takes a cheaper path: it freezes the original weights and trains a small set of added matrices instead. The original paper reports cutting trainable parameters by up to ten thousand times and GPU memory by three times while keeping quality comparable to full fine-tuning.[1] That gap is why LoRA is the default for most teams, and full fine-tuning is reserved for when you genuinely need to shift the whole model.

The harder question is whether to fine-tune at all. Fine-tuning changes how a model behaves, a format it should always produce, a tone, a narrow task it keeps fumbling, but it is a poor way to give a model facts, because those facts get frozen into the weights and go stale. When the need is current or private information, retrieval is the better tool: RAG fetches the relevant text at question time instead of baking it in.[2] A good rule of order is to try prompting first, then RAG, and reach for fine-tuning only when those fall short, since each step up costs more to build and maintain than the one before it.

References

  1. LoRA: Low-Rank Adaptation of Large Language Models (Hu et al., 2021) — arXiv
  2. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (Lewis et al., 2020) — arXiv

Common questions

When should I fine-tune instead of prompting or using RAG?
Fine-tune to change how a model behaves: a consistent format, a tone, a narrow task it keeps getting wrong despite clear instructions. To give it new or changing facts, use RAG instead, since fine-tuned facts are baked in and go stale. Try prompting first, RAG second, fine-tuning last, because each costs more than the one before.
What is the difference between LoRA and full fine-tuning?
Full fine-tuning updates all of a model's weights, which is thorough but expensive in compute and memory. LoRA freezes the original weights and trains a small set of added matrices, cutting trainable parameters by orders of magnitude while keeping quality close. LoRA is the default for most teams; full fine-tuning is for when you need to move the whole model.

Related