LoRA Explained in 2 Minutes: Fine-Tune LLMs Without Burning Your GPU

Описание: How can you fine-tune a large language model without updating billions of parameters or setting your GPU’s VRAM on fire?

In this two-minute video, we explain LoRA (Low-Rank Adaptation) through a practical and slightly absurd example: creating a custom LLM persona.

The video covers:
• Why full fine-tuning is computationally expensive
• How LoRA represents the weight update as two low-rank matrices
• Why only matrices A and B need to be trained
• How LoRA reduces the number of trainable parameters
• How the adapted model processes an input during inference
• The importance of the rank parameter, r

This video was created for the Natural Language course, as part of the Master’s Degree in Computer Science and Engineering at Instituto Superior Técnico (IST), Portugal.

#LoRA #LowRankAdaptation #LLM #FineTuning #MachineLearning #DeepLearning #NaturalLanguageProcessing #ArtificialIntelligence #IST