LoRA

AI Glossary

LoRA is a cheap way to fine-tune a giant model by training only a small adapter on top — minutes instead of days.

What it really means

LoRA stands for Low-Rank Adaptation. It’s a technique that lets you take a big, pre-trained AI model — the kind that costs millions to build — and teach it a new trick without retraining the whole thing. Think of it like buying a high-end pickup truck and then swapping out just the suspension for off-roading, instead of rebuilding the entire truck from scratch.

Here’s the technical bit, in plain English: Large AI models have thousands of internal “knobs” that get tuned during training. Normally, if you want the model to learn something new, you have to retune all of those knobs — which takes days or weeks and costs a fortune in computing power. LoRA works by adding a tiny set of new knobs (called a “low-rank adapter”) that sits on top of the original model. You only train those new knobs. The original model stays frozen. The result is a specialized model that’s maybe 1% the size of the original, but can do the new task almost as well.

I help businesses in Orlando use LoRA because it makes custom AI practical for companies that don’t have a data science team or a six-figure cloud budget.

Where it shows up

You’ll mostly see LoRA used with large language models — things like GPT, Llama, or Mistral. But it also works with image generation models like Stable Diffusion. If you’ve ever seen someone generate “a cat in the style of Van Gogh” with a custom look, that’s likely LoRA at work.

In the business world, LoRA is the reason you can take a general-purpose AI assistant and teach it your company’s specific jargon, your customer service scripts, or your legal document templates — without needing to rent a data center.

It’s also common in open-source AI communities. Platforms like Hugging Face host thousands of LoRA adapters that people share for free. You can download one that’s already trained for medical text, legal writing, or even restaurant menu descriptions, and then tweak it further for your own use.

Common SMB use cases

• A Winter Park dental practice fine-tunes a language model on their patient intake forms and common dental procedure descriptions. The LoRA adapter lets the AI answer patient questions about root canals or billing with the exact language the practice uses, not generic textbook answers.
• A Maitland HVAC company trains a LoRA adapter on their service manuals and common repair logs. Their field technicians can then ask the AI “What’s the fix for a buzzing condenser on a 2022 Trane?” and get an answer that matches their specific equipment and procedures.
• A downtown Orlando law firm uses LoRA to adapt a legal writing model to Florida-specific statutes and their own contract templates. The AI drafts initial versions of non-disclosure agreements or lease addendums that follow the firm’s preferred language — cutting drafting time from hours to minutes.
• A Lake Nona restaurant trains a LoRA adapter on their menu descriptions and customer reviews. The AI can then generate new menu item descriptions, social media posts, or email newsletters that match the restaurant’s voice and highlight their signature dishes.
• A Clermont pool service company fine-tunes a model on their maintenance checklists and common customer questions. The LoRA adapter helps their customer service team answer “Why is my pool turning green?” with the same troubleshooting steps the technicians use in the field.

The common thread: LoRA makes custom AI affordable for businesses that have maybe 50 to 200 example documents, not millions.

Pitfalls (what gets oversold)

The biggest oversell I see is the idea that LoRA can fix a fundamentally bad base model. If the underlying model is weak — say, it can’t handle basic reasoning or has terrible factual accuracy — adding a LoRA adapter won’t magically fix that. You’re still building on a shaky foundation.

Another common mistake: people think LoRA training is zero-effort. It’s not. You still need clean, well-organized data. A LoRA adapter trained on messy or contradictory examples will produce messy or contradictory results. I’ve seen businesses dump a pile of random emails into a training set and expect magic. That doesn’t work.

There’s also the “one adapter to rule them all” trap. LoRA adapters are specialized. If you train one for legal document drafting, it won’t also be good at writing marketing copy. You need separate adapters for separate tasks. That’s fine — they’re tiny files — but you need to keep them organized.

Finally, LoRA doesn’t give you the same quality as full fine-tuning for very complex, nuanced tasks. If you need a model that deeply understands a narrow, highly technical field — like interpreting medical imaging reports — full fine-tuning might still be necessary. LoRA is a practical shortcut, not a universal replacement.

Related terms

• Fine-tuning — The general process of taking a pre-trained model and training it further on a specific dataset. LoRA is one method of fine-tuning, but there are others (like full fine-tuning or adapter layers).
• RAG (Retrieval-Augmented Generation) — A different approach where you don’t train the model at all. Instead, you give it access to a database of your documents, and it retrieves relevant info on the fly. RAG is often easier to set up than LoRA, but can be slower and less precise for very specific tasks.
• Adapter — A small, trainable module that’s inserted into a larger model. LoRA is a specific type of adapter that uses low-rank matrices to keep the number of new parameters tiny.
• Base model — The original, pre-trained AI model that you start from before applying LoRA or any other fine-tuning method. The quality of your LoRA adapter depends heavily on the quality of the base model.
• Inference — The process of running a trained model to get an output. Once you’ve trained a LoRA adapter, you load it alongside the base model during inference to get your specialized results.