PEFT (parameter-efficient fine-tuning)

PEFT (parameter-efficient fine-tuning) is a group of methods for training an existing AI model further on your own examples with far less computing power. Only a small add-on or a small part of the model is trained, and the rest stays frozen. Each result is a small file, so a company can keep one model with a separate add-on per task, such as sorting emails or drafting replies.

How does it work?

Full fine-tuning adjusts every weight of a model, and training needs much more memory than the weights alone: for each trained weight it also stores a gradient and, with the Adam optimiser, two running averages. According to Lialin et al. (2023), training in practice takes 12 to 20 times more GPU memory than the model weights. PEFT methods freeze most weights and train only a small number of parameters, so the gradients and optimiser states shrink with them. Hugging Face's PEFT library says these methods perform comparably to full fine-tuning at much lower computing and storage costs.

Which methods are there?

Lialin et al. sort the methods into three groups, plus hybrids that combine them:

• Adding parameters. Adapters (Houlsby et al., 2019) insert small trainable layers into every Transformer block. On the GLUE benchmark they came within 0.4% of full fine-tuning while adding 3.6% parameters per task. Soft prompts instead train a few extra vectors that the model reads as virtual tokens in front of the input: prefix tuning (Li and Liang, 2021) with 0.1% of the parameters, and prompt tuning (Lester et al., 2021), which matched full fine-tuning once models passed billions of parameters.
• Selecting parameters. BitFit (Ben-Zaken et al., 2021) trains only the bias terms, and on small to medium datasets it was competitive with fully fine-tuning BERT.
• Reparametrising. LoRA (Hu et al., 2021) writes the change to a weight matrix as the product of two thin matrices; for GPT-3 175B it trained 10,000 times fewer parameters than full fine-tuning. QLoRA (Dettmers et al., 2023) trains LoRA on a model compressed to 4 bits with quantization: fine-tuning a 65-billion-parameter model then needs less than 48 GB of GPU memory instead of more than 780 GB.

Lialin et al. note that reparametrisation methods such as LoRA have become widely popular. In a comparison of nine post-training libraries in Hugging Face's TRL v1.0 announcement (March 2026), eight supported LoRA and five also QLoRA.

Where does it go wrong?

• The model learns less. Biderman et al. (2024) found that LoRA learned substantially less than full fine-tuning on programming and mathematics, and forgot less of what the base model could already do.
• Some methods slow the model down. Adapters add layers that run one after another: on GPT-2 medium with one request at a time, they made a forward pass 21 to 30% slower (Hu et al.). LoRA's matrices can be added into the original weights after training, so it adds no delay.
• Soft prompts take up room. A prefix uses part of the input length the task could otherwise use, and Hu et al. found prefix tuning hard to optimise.
• Reported savings can mislead. Lialin et al. found clear gaps between reported and actual numbers of trainable parameters, and methods that had beaten LoRA in earlier papers struggled to do so with limited hyperparameter tuning and few training epochs.

Example

Suppose a customer service desk wants one model that sorts incoming emails and one that drafts replies in the company's tone. With full fine-tuning you would store two complete copies of the model. With LoRA you train two small adapters on the same base model, load the right one for each task and test both on emails the adapters have not seen.