Term
PEFT (Parameter-Efficient Fine-Tuning)
PEFT bundles methods that adapt a pretrained LLM by training only a small fraction of its weights — typically under 1%, instead of all billions of parameters.
PEFT — explained in more detail
Full fine-tuning of a modern LLM is expensive: all billions of parameters must sit in GPU memory, gradients and optimiser states multiply the footprint, and every use case produces its own full model copy. PEFT sidesteps this by freezing the original weights and training only a small number of additional or selected parameters. The result: training memory often shrinks tenfold or more, the trained adapters are a few megabytes in size, and several use cases can run on the same base model with different adapters in parallel. Hugging Face’s peft library exposes most common methods as one-liners.
Example / Practical context
Classic case: a Llama or Mistral base model should answer customer support mails in the company’s house tone. Full fine-tuning would tie up several high-end GPUs for days. With LoRA — the most popular PEFT variant — a single A100 GPU for hours is often enough; the trained adapter is a few hundred megabytes and can be plugged into or detached from the base model at runtime.
Distinction from related concepts
LoRA (low-rank adaptation) is the best-known concrete PEFT method — it injects small rank-r matrices into the attention layers. QLoRA combines LoRA with 4-bit quantisation of the base model and cuts memory further. Full fine-tuning trains all parameters, is more expensive and produces a complete new model copy. Prompt tuning and prefix tuning are further PEFT methods that inject learned embeddings ahead of the input rather than touching weights.
Entdecke mehr
AI Workflows by Keyword: How We Make Recurring Routines Enforceable
A typed keyword triggers a fixed AI routine — and every single step must be committed before the next one appears. Why that's the actual trick.
GlossarDPO (Direct Preference Optimization)
DPO is a fine-tuning method that aligns language models directly to human preference pairs — without a separate reward model and without reinforcement learning.
LexikonFine-Tuning Explained — When, How, With What
When fine-tuning pays off, which methods exist (SFT, DPO, LoRA, QLoRA), and what AMD vs. NVIDIA hardware actually means in practice.