How to fine-tune a model using LoRA (step by step)

LoRA is a genius idea.

By the end of this video, you'll know everything important about how LoRA works. I'll show you an example where we'll fine-tune a model with LoRA using two different datasets. I'll later load the model and plug the correct adapter to solve different classification tasks.

I teach a live, interactive program that'll help you build production-ready Machine Learning systems from the ground up. Check it out here:

To keep up with my content:

[bithigh]

Thanks Underfitted!! Guys let's comment, like, and subscribe. This channel is pure gold!

[solidkundi]

love it... as i am still a noob..would love to see a llm example with summarization model ..and to see the format involved .. thank you again!!!

[flcor]

One note: the row-column decomposition is valid for matrices whose rows (columns) are not linearly independent — that’s probably why they train on the row-columns themselves and not on general matrices that cannot be factorized into row-column form. So, there’s clearly a tradeoff here between memory and linear independence.

[kaloyanmirchev]

Thank you for presenting such great ideas. My imagination surely goes wild when I attempt to think of possible applications..

[UtopIA-IA]

OMG! that is so powerful, thank you, I am alone doing projects of this type and this will be very useful for me, thanks for sharing you knowledge.

[jayng]

Nice tutorial, would like to ask how to fine tune an AI model that generate interior design?

[sirojiddinnuriyev]

There are a lot of contents how to fine-tune LLMs with LoRA or QLoRA. You gave us same food just with ‘apple genius’ keyword.

[monugarg]

Hey sir. Very good explanation. Sir is it possible for you to make a video on Ai Agents and tools please.

[mabadolat]

This is great stuff, Santiago! I wish you had posted this video a few weeks ago. We just completed our final class project where we trained five different BertClassifier models on five different tasks. Our fine-tuning and inference code structure is very similar to yours. We definitely could have used this approach to use just the specialized adapters instead of the full BERT models.

However, I have one question: I'm not clear whether the full model will ever be used during this process after we get fine-tuned adapters, or just the fine-tuned weight matrix for evaluation and inference?

[tecbrain]

Fantástico vídeo. Gracias por el tiempo que inviertes. Ahora me queda entender bien el código.

[syedasadzaman]

Good Stuff! Santaigo.

The channel name should be `Tutorials That do not Suck!` =}

[allanmogley]

How do you make those models that interact with data,
Like I once saw someone create something really amazing that inteprets data from a a database and makes interpretations and reports from the data wothout hallucinating (It only fetches from the underlying DB)

[krishnasoni]

I cannot talk with the agent, the connection is established but it aint respond or neither taking image i/p please suggest something

[AngusLou]

Can you make a Google Colab notebook for the same fine-tuning?

[devevangelista]

if run this with 16G memory and RTX 2060 could work?

[m.active]

there isn't a general solution to decompose a matrix of M*N into two vectors of M*1 and 1*N. If that was the case we could have some all the issues in the data compression by now. A lossless compression of 99.99% for huge matrix is a strange achievement.

[kencottrell]

When you have the original model + plus the Adapter model, can the original model still solve the save generic tasks? In other words, can you perform original inferencing tasks PLUS specific tasks?

[robertobreve]

The jupyter notebook has broken images.

[philyou]

Missing all the info that is needed to implement the idea on own data set