The Evolution of Fine-Tuning

Fine-tuning of transformers hasn’t always been the same. It’s has significantly simplified with the latest LLMs like Llama at the expense of lesser control reducing the number of potential use cases. It isn’t necessarily bad cause it’s a process of evolution and adoption but you might want to understand what you could achieve with smaller less-expensive models for which I’m doing a deep fine-tuning walkthrough for you.
This is a second part of my report about researching transformers on a Grammar task where I’m explaining a bit of a theory behind fine-tuning and various optimization techniques. I’m also giving you a means of addressing fine-tuning architecturally highlighting the points of consideration before you want to get into this game.
Source code (evaluation, fine-tuning, analysis, infra): github.com/iliazlobin/transfo...
Timelines
00:00:00 Introduction
00:00:14 What we are going to talk about
00:02:23 What is Fine-Tuning really is?
00:04:10 Why should you Fine-Tune?
00:06:02 When to restrain from Fine-Tuning?
00:07:52 Challenge yourself with this questions to find if Fine-Tuning is right for you
00:09:39 How Instruction Following of the 1.3B parameter model has improved over the 175B model?
00:11:29 Prerequisites for Fine-Tuning
00:15:01 How to properly approach Evaluation: Frameworks, Scenarios, Metrics
00:18:17 Understanding Perplexity
00:19:59 A few words about online evaluators
00:20:50 Optimization methods for limited hardware resources (quantization, PEFT, soft prompts)
00:23:04 Taking the first look into the notebooks
00:24:14 What are the models that we are going to be fine-tuning today?
00:26:27 *Fine-tuning of the base T5 model: complete walkthrough*
00:50:38 Fine-tuning gpt2 decoder-only transformer
01:02:44 Fine-tuning llama-2 7b with bitsandbytes and Lora optimizations
01:09:38 Axolotl can do many of those fine-tuning tricks for you, but not all
01:11:58 Fine-Tuning LLM for Magic the Gathering Draft recommendations
01:16:11 Conclusion
About
CZcams: / @iliazlobin
GitHub: github.com/iliazlobin
WebSite: iliazlobin.com/blog
LinkedIn: / iliazlobin
X: x.com/iliazlobin
Email: [[e-mail]](mailto:[e-mail])
As always, feel free to reach out on social media or ping me if you’re around NYC and would like to meet and chat.
Hope you’ve enjoyed it and learned something new!
Thanks for watching!

[@KenRaps]

This is really in-depth. I enjoyed watching it. Thanks, Ilia!

[@iliazlobin]

I really appreciate it! Would love to know if all aspect are clear, or you'd like something to be explained further

[@Archiiee1]

Hi from Argentina. I will begin my journey on fine tuning with you 😊

[@iliazlobin]

Thank you so much! You could also check the repo, there's much more details. Let me know if you've any questions