zero2robot: a free from-scratch course whose LoRA chapter is a guided reading of PEFT

[kaushikb11]

Sharing this since PEFT is the "read the real thing" reference for the LoRA chapter of the course.

zero2robot is a free, from-scratch course on embodied AI: you build a robot brain one readable file per chapter, from a bare MuJoCo loop up to a vision-language-action policy you train and drive in the browser.

In the fine-tuning chapter (ch5.6) learners implement LoRA from scratch, one readable file, and then read the PEFT implementation as the production ground truth to compare against: where the adapters attach, how merging works, why it saves memory. Everything runs on a free Colab T4 or a CPU laptop.

Textbook: https://www.zero2robot.com/
Source: https://github.com/kaushikb11/zero2robot

Thanks for PEFT; having a clean, real LoRA to point learners at after they build their own is exactly the "read the real thing" moment the course is built around. Feedback welcome.

[githubnemo]

Hey, thanks for the heads-up.

Personally I would have expected a bit less fluffy text and maybe a few more visualizations on how the architecture of LoRA works. Just having the code makes it a bit dense.

I liked the forgetting vs. learning visualization but it was not clear to me what -1 for task_A meant, is it a relative performance loss? Is it absolute values?

In any case it might be worthwhile to have some pointers to methods that target forgetting such as LoRA intruder dimension reduction or using Orthogonal Subspace Fine-Tuning for continuous learning in the context of robotics.