This lab works on phones, but writing code is much easier on a laptop.
LLM · Capstone: Build and Evaluate Your Own LLM System · LM12-L4 · Advanced
The last milestone is writing, and the most common failure of a technical report is a claim that is bigger than its evidence. A reviewer asks the same questions every time: compared with what, how many runs, how uncertain, and was the test set touched?
You will play reviewer on six claims, turn the reviewer’s questions into a checker, and implement the paired test that settles “is A really better than B”.
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 applies a rubric by hand to six claims. Stage 2 encodes the rubric as code and runs it on hidden claims. Stage 3 implements an exact paired sign test on per-example results. No model is run: the reports and per-example results are fixed data. Your code is graded on hidden claims and result lists.
Educational simplification: a rubric with five checks and a sign test on pass/fail results. Real reviews also look at data quality, metric choice and the plausibility of the story.
Get notified when new robotics labs are released. One email per new lab, nothing else.