PEFT

You are about to learn one of the most practical ideas in modern AI: how to take a huge model that someone else spent months and a fortune training, and adapt it to your own task — without retraining it from scratch. Understanding PEFT (Parameter-Efficient Fine-Tuning) will let you fine-tune a model on your own laptop or in your own company, even when the model has billions of parameters. This unlocks the Fine-Tuning parent concept and all its children: LoRA and QLoRA (the most popular PEFT methods), Instruction Tuning (making a model follow commands), Catastrophic Forgetting (the risk you are avoiding by using PEFT), and Multi-Tenant Adapter Serving (using one base model for many customers). With PEFT, you will be able to build systems that are cheap, fast, and private — and you will understand what every AI engineer means when they talk about 'adapters' and 'frozen weights'.

Imagine you are a manager in a large company. The company has a huge, well-trained workforce — thousands of people who can do many tasks. Now the company gets a new project: handling customer complaints in a new, very technical domain (say, legal contracts). Training the entire workforce from scratch would take years and cost a fortune. Instead, you try something smarter: you keep the existing workforce, but you give each team a small 'adjustment manual' — a few pages of instructions that tell them how to apply their general skills to this specific domain. The manual is tiny compared to the full training, but it changes the output just enough. This manual is like a PEFT adapter: a small set of new instructions (parameters) that you add to a frozen, already-trained model. You train only these new parameters, not the whole model. The model keeps all its original knowledge — it still knows English, grammar, and general facts — but it also learns your specific domain. And because the manual is small, you can swap it out, share it, or store many different manuals for different tasks. That is PEFT in a nutshell.

Think of a chef who has trained at a famous culinary school. The chef knows a thousand recipes — French, Italian, Indian, etc. Now a new restaurant opens: it specialises in Gujarati thali. Retraining the chef from scratch would mean sending them back to school for years. Instead, the restaurant gives the chef a small recipe card — just a few key adjustments: use this particular spice blend, this style of dal, this proportion of vegetables. The chef still knows how to chop, sauté, and plate — all the basics — but the recipe card directs their skills to the new cuisine. Now, if the restaurant later wants to serve Punjabi food, they just give the chef a different recipe card. The chef's core knowledge stays the same; only the small set of instructions changes. This is exactly how PEFT works: you take a pretrained model (the chef), freeze its weights (the chef's core knowledge), and add a tiny set of new parameters (the recipe card) that you train on your specific task. The new parameters are so small that they can be versioned, stored, and swapped like configuration files. And here is where the analogy stops: a human chef can only hold one recipe card in mind at a time, but a model with PEFT can have many adapters loaded dynamically — thousands of different 'recipe cards' can be attached to the same base model, and you can switch between them in milliseconds.

Parameter-efficient fine-tuning (PEFT) is the family of methods that adapt a pretrained model to a new task by training a small number of added or selected parameters, while freezing the original (base) weights to preserve the model's general capabilities.

You are at the start of the Fine-Tuning branch. Fine-tuning is the general practice of taking a pretrained model and adjusting it to a new task. PEFT is the most efficient way to do that. Your library notes mention that LoRA and QLoRA are the most popular PEFT methods, and Multi-Tenant Adapter Serving is what makes PEFT economically powerful: one frozen base model can serve many customers, each with their own small adapter. PEFT also directly addresses the Catastrophic Forgetting risk — because you are not changing the base weights, you do not lose the original abilities. Later concepts like Instruction Tuning and Classification Fine-Tuning can be done with PEFT methods.