This lab works on phones, but writing code is much easier on a laptop.
LLM · Efficient Fine-Tuning and Model Adaptation · LM8-L2 · Advanced
LoRA rests on one observation and one identity. The observation: the change a task needs in a weight matrix is often well captured by a few directions. The identity: a product of two thin matrices is a matrix of low rank, and can be added to the weight after training.
You will find how much rank a constructed update needs, write the forward pass, and merge the adapter into the weights.
You have passed every stage. Now connect it back to the theory and move on:
Plain Python with math (and slamkit for small matrix helpers: matmul, transpose, solve, inverse, identity). Images, point clouds and detections arrive as lists inside the dictionary D; an image is a list of rows, each row a list of pixels. There is no NumPy or OpenCV: you write the algorithm yourself, which is the point of the lab. Print with print().
Stage 1 explores the trade-off between the rank and the part of an update that can be represented. Stage 2 computes the LoRA forward pass. Stage 3 merges the adapter and counts parameters. The matrices are small and constructed; no model is trained. Your code is graded on hidden matrices.
Educational simplification: a 16 × 16 update with five known singular values, matrices of a few rows and columns, plain Python lists. Real LoRA layers have thousands of rows and columns and are trained by gradient descent; the update of a trained model is not known in advance. All data in this lab are invented.
Get notified when new robotics labs are released. One email per new lab, nothing else.