Encyclopedia · 311 concepts
Learn this concept as a feed

NLP & Language · intermediate · concept 129 of 311 · Updated Nov 4, 2026

LoRA & Parameter-Efficient Fine-Tuning

Instead of updating billions of weights, LoRA freezes the model and trains two small low-rank matrices per layer, often under one percent of the parameters. Fine-tuning that needed a cluster now fits on one consumer GPU, and swapping adapters swaps skills without touching the base model.

L7 · LLM and generative AI engineerDevelopersTech leadersQA and testing

Key terms

Low-rank adaptationAdaptersQLoRAFrozen base modelPEFT

Learn these first

Where you meet it in the real world

Custom-styled image models, domain-tuned chatbots, per-customer adapters served from one base

Videos

Guides and articles

Courses, papers, and more

This unlocks