Standard 9 stops to get here · leads to 2

Parameter-Efficient Fine-Tuning (PEFT)

A family of methods that adapt a large pretrained model by training a small number of new or selected parameters while leaving the original weights frozen.

Your route here

9 stops · basics first
  1. Machine Learning ✓ understood

    Building systems that learn patterns from data instead of following hand-written rules, getting better at a task as they see more examples.

  2. Neural Network ✓ understood

    A model built from layers of simple units, each taking a weighted sum of its inputs and applying a nonlinearity; training adjusts the weights so the whole stack maps inputs to useful outputs.

  3. Parameter ✓ understood

    A number inside a model, a weight or a bias, that training adjusts to reduce the loss. Together, a model's parameters are everything it has learned.

  4. Dataset ✓ understood

    A collection of data examples used for training, validating, or testing machine learning models.

  5. Training ✓ understood

    The process of fitting a model to data by repeatedly measuring how wrong its outputs are and adjusting its parameters to reduce that error.

  6. Unsupervised Learning ✓ understood

    Learning from unlabeled data to discover hidden patterns, structures, or relationships without explicit target outputs.

  7. Self-Supervised Learning ✓ understood

    Training on unlabeled data by making the data supply its own answers: hide or alter part of an input and have the model predict or match it.

  8. Pre-training ✓ understood

    Training a model on a large dataset (often self-supervised) before fine-tuning on specific tasks, enabling transfer learning.

  9. Fine-Tuning ✓ understood

    The process of further training a pre-trained model on a specific dataset to adapt it for a particular task or domain.

  10. Parameter-Efficient Fine-Tuning (PEFT) · you are here ✓ understood

Picture it

Full fine-tuning

  • Every weight is trained
  • Optimizer state for every weight
  • One full model copy per task

PEFT

  • Base model frozen
  • Train a small add-on (adapter, LoRA)
  • Swap small add-ons per task
Notice what gets saved per task: a whole model on the left, a small set of extra weights on the right that sits on one shared base.

Parameter-efficient fine-tuning (PEFT) adapts a large pretrained model to a new task without retraining all of its weights. Ordinary fine-tuning updates every parameter, which for a model with billions of them means a lot of GPU memory during training and a full-size copy of the model for every task. PEFT methods freeze the original model and train only a small set of parameters, a tiny fraction of the total. For GPT-3, LoRA cut the trainable parameters by a factor of 10,000.

The main approaches

  • Adapters insert small trainable layers between a model’s existing layers. Houlsby et al. introduced them in 2019 and came within 0.4% of full fine-tuning on the GLUE benchmark while adding about 3.6% extra parameters per task.
  • LoRA learns a low-rank update to existing weight matrices. The update can be merged back into the weights, so it adds no extra latency at inference. It is the most widely used PEFT method; see the LoRA paper.
  • Prompt tuning and prefix tuning learn a few trainable vectors that are prepended to the input, leaving the whole network untouched.

Why it matters

Training fewer parameters means less memory for gradients and optimizer state, and pairing PEFT with quantization of the frozen base shrinks the rest. Together they put fine-tuning of large models within reach of a single, even consumer-grade, GPU. And because the base model is shared, one deployment can serve many tasks by swapping small add-ons, each cheap to store and share.

The catch

A small add-on has limited capacity. PEFT works best when the task is close to what the base model can already do, such as a new format, style or domain; teaching a model something far outside its pre-training is harder to squeeze into so few extra weights.

Where it sits

Before this

Fine-Tuning Parameter
Parameter-Efficient Fine-Tuning (PEFT)

Leads to

LoRA Prompt Tuning

Explore nearby

Training LoRA Low-Rank Adaptation - a parameter-efficient fine-tuning method that updates only small low-rank matrices instead of full weights. Language & LLMs Prompt Tuning Learning continuous prompt embeddings while keeping the LLM frozen, an efficient alternative to fine-tuning. Language & LLMs Instruction Tuning Fine-tuning LLMs on diverse instruction-following tasks to improve zero-shot performance on new instructions. Training Transfer Learning Leveraging knowledge learned from one task/domain to improve performance on a related task with less data. Shipping AI Quantization Reducing model precision (FP32 → INT8) to decrease size and increase inference speed with minimal accuracy loss.

Sources

  1. Neil Houlsby et al., "Parameter-Efficient Transfer Learning for NLP" . ICML 2019; adapter modules
  2. Edward J. Hu et al., "LoRA: Low-Rank Adaptation of Large Language Models" . ICLR 2022
  3. Hugging Face PEFT documentation . Library documentation