This lab works on phones, but writing code is much easier on a laptop.
LLM · Fine-Tuning for Classification · LM6-L1 · Beginner
A classifier that says “no hazard” every time is right 92% of the time on a dataset where 8% of the messages are hazards. Accuracy alone would rank it as excellent.
You will pick a threshold with that trap in mind, write the metrics that look at the rare class, and then hunt for the other common reason that a good score is not trustworthy: a feature that works only by accident.
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 threshold of a score-based classifier on imbalanced data. Stage 2 computes the confusion matrix and the four main scores. Stage 3 finds a shortcut feature by comparing training and validation behaviour. The scores and examples are generated numbers; no language model is used. Your code is graded on hidden data.
Educational simplification: binary labels, one score per message, a handful of binary features, small tables of examples. Fine-tuned language models are evaluated in the same way, with the same metrics, on text examples. All data in this lab are invented.
Get notified when new robotics labs are released. One email per new lab, nothing else.