Low-Rank Adaptation (LoRA): From Intuition to Implementation to Interview Questions

Author(s): Harsh Maheshwari

Originally published on Towards AI.

Delving Deeper into LoRA for LLMs
Photo by Digital Content Writers India on Unsplash

LLM are getting larger and larger each day, recently, Meta has announced that they are currently working on the 400b Llama 3 model. It has been really awesome to have these larger models that have more intelligence (in general) and can work on a variety of tasks. But there have been times when, for a particular use case/dataset, these off-the-shelf LLMs have not worked out as well as smaller trained models. So the next step that comes to mind of ML practitioners is to fine-tune the LLM on our dataset and check the performance. Now, here, the “larger” LLM comes back to haunt us, since training such models requires extensive hardware. Training the full model is compute and memory-intensive as well as data-intensive since they might easily overfit on our small fine-tuning dataset. Forget about the larger LLMs (order of 70B). Even smaller models with 7B parameters (not really small, though) are difficult to fine-tune because of the above constraints. This is when LoRA comes into play.

LoRA is a machine learning technique used for fine-tuning a Large Language Model (LLM) on a small dataset by adding a few sets of new parameters in… Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI