AI Glossary 10 - Fine-Tuning and Adaptation

No mechanism post on this site teaches fine-tuning end to end yet; the nearest neighbor is how vertical small language models work, where adaptation is the whole business case. For the bigger picture of tuned versus prompted behavior, our LLM concepts deep dive rounds it out. This tenth glossary lesson therefore carries the vocabulary itself: the fine-tuning terms that separate a tuned model from a prompted one, from transfer learning to LoRA to catastrophic forgetting. Each entry gets a sentence or two, with the choices and tradeoffs stated plainly.

The adaptation ladder

Three terms for what it means to keep training a finished model.

• Fine-tuning: Continuing to train an existing model on new data, adjusting the weights so it behaves differently in a chosen domain or style. Prompting changes what you say to the model; fine-tuning changes what the model is, and the results persist in every later conversation.
• Transfer learning: The broader principle underneath: knowledge learned on one task is reused as the starting point for another. Fine-tuning a pretrained language model is the most common instance; the pretraining already paid for the grammar, facts, and fluency that your small dataset could never teach.
• Full fine-tuning: Updating every weight in the model during adaptation, as opposed to the budget methods later in this lesson. It offers the largest behavior change per epoch and demands the most hardware, the most data, and the most care, since everything the model knew is now negotiable.

Recipes with names

Four fine-tuning terms for the dataset shapes you will hear offered as services.

• Supervised fine-tuning: Training on input and output pairs, the plain labeled-data recipe, where the gradient follows a known correct answer. It is the workhorse behind most domain adaptation and the first stage of most chat-model training pipelines.
• Instruction tuning: Supervised fine-tuning on examples phrased as instructions and responses, so the model learns the form of being told what to do. The distinction from plain SFT is the data, not the math: the pairs look like requests a user would actually send.
• Domain adaptation: Tuning toward one field’s vocabulary and conventions, legal language, clinical notes, internal ticket phrasing, usually to lift reliability on inputs that a general model handles superficially. Vertical model builders treat this as their core product step.
• Multi-task learning: Tuning a single model on several task mixtures at once so the tasks regularize each other, rather than fitting a separate specialist per task. The model shares what generalizes and the operator maintains one deployment instead of five.

Adaptation on a budget

Four terms for the methods that train a tiny fraction of the parameters and leave the rest frozen.

• Parameter-efficient fine-tuning: The family name for all such methods, abbreviated PEFT, covering adapter layers, low-rank updates, and soft-prompt variants. The pitch is constant: store megabytes instead of gigabytes per adaptation, and often run several adapted variants of one base model on the same serving footprint.
• LoRA: Low-rank adaptation, the method that made PEFT mainstream: freeze the original weight matrices and learn small low-rank update matrices beside them. The updates merge back into the weights for serving; QLoRA pairs the trick with a 4-bit quantized base model so a single consumer GPU can tune one, and DoRA and the variant family refine the decomposition further.
• Adapters: Small trainable modules inserted between the frozen model’s layers, the original PEFT design. Each new task trains a new set of modules and the base model never moves, which keeps a library of task variants cheap to keep and safe to swap.
• Soft prompts: Training a short sequence of continuous vectors prepended to the input, prompt tuning and prefix tuning being the two named variants. Unlike a written prompt, these vectors are learned by gradient descent and never readable, but they occupy the same position in the context and cost the same per request.

What can go wrong

Two terms for the failure modes every tuning run must manage.

• Catastrophic forgetting: The tuned model losing previously held skills as it absorbs the new ones, fluent in the new domain and newly clumsy at everything else. Small learning rates, mixed-in general data, and PEFT methods that leave the base weights untouched are the standard countermeasures.
• Layer freezing: Locking chosen layers, usually the early ones that encode general features, so tuning touches only the rest. It is the cheapest defense against forgetting and against overfitting a small dataset, at the cost of capping how much the adaptation can change.

Common confusions

Three pairings that decide budgets when they are confused.

• Instruction tuning vs supervised fine-tuning: SFT is the training procedure; instruction tuning is SFT with instruction-shaped data. Every instruction-tuned model was supervised fine-tuned, so the words differ in what they tell you about the dataset, not the optimizer.
• LoRA vs full fine-tuning: LoRA changes behavior through small low-rank updates and preserves the base model; full tuning rewrites all the weights and can reach changes LoRA cannot. Teams regularly start with LoRA because a failed adapter is discarded, while a failed full tune has to be restored from backup.
• Prompt tuning vs prompting: Prompting writes text anyone can read and edit; prompt tuning learns vectors by gradient descent that no one can inspect. The names sound adjacent and the practices sit far apart: effort, cost, and who is allowed to change the result all differ.

Further reading

Two papers that define the budget methods above:

• LoRA: Low-Rank Adaptation of Large Language Models, the paper that made adapter-scale training a default choice.
• QLoRA: Efficient Finetuning of Quantized LLMs, the combination of 4-bit quantization and LoRA that moved serious tuning onto single GPUs.

This is lesson 10 of the AI glossary. Lesson 09 covered the inference and serving terms that any tuned model still has to run under, and quantization reappears here inside QLoRA from the other direction. The next lesson follows adaptation toward its most contested endpoint: alignment, RLHF, and the reinforcement learning vocabulary behind them, and the vertical models explainer shows the business side of these choices in production.