Instantiation of generate-exam.md for the three-week offering, which has one combined exam and no midterm. It has typically run in September (Paris), but the recipe depends on the offering’s length, not the season. Read generate-exam.md first for the LaTeX conventions and workflow, then apply the parameters below. The nine-week offering uses generate-midterm.md and generate-final.md instead.

Students: run this yourself to make practice exams. You will not get the same questions the instructor got, but you will get questions built from the same inputs.

Parameters

Parameter	Value
Exam type	combined (single exam; no midterm)
Agenda	docs/agenda/<term>.md for the term (e.g. 2026-summer.md), every meeting up to the exam. Meeting numbers and content vary by year
Term page	docs/terms/<term>.md: exam date, lectures skipped or merged, and any bonus lectures given after the exam was set. Everything it lists as skipped or not on the exam is out of scope
Assignments	Every assignment and project milestone due before the exam (in 2026: Assignment 1 and the project proposal); at least one question on each
Hands-ons	The ones the agenda records for the term, not a fixed list
Past exams to imitate	The most recent combined exam(s) (docs/combined/<YYYY>/), plus the most recent nine-week midterm and final: a combined exam mixes both
Length	6 pages (7 if the feedback item spills), 75 points, designed for 45–50 minutes with the full class period available (the nine-week exams are 4 pages / 50 points)
Question mix	Mostly select-all / multiple choice and yes/no with “Why or why not?”; keep pure short answers to a handful (graded by keyword); where a question has two parts, make one part multiple choice and one a box; end with a 2-point feedback section
Instructor output	/Users/feamster/Dropbox/Tmp/exams/ml-systems/combined/<YYYY>/ (not committed until after the exam)

Coverage

Proportional to time spent in that term’s agenda, and nothing outside it. A three-week term moves through the whole course, so expect roughly: why ML for networks and the pipeline; the security, performance, and resource-allocation applications; data acquisition and representation; training and evaluation; supervised models; deep learning; unsupervised learning; and generative models. Treat that list as orientation only. Drop whatever the term page says was skipped, and anything the agenda says will not be asked.

Flagged questions

Cover every question the agenda flags as a good exam question, including any the instructor adds in an exam-review meeting near the end of the term. Collect them before writing anything.

Standing ground rules

Set in the 2026 review and kept unless that term’s agenda says otherwise:

• No definition regurgitation (e.g. Mitchell’s definition), no logistics or history, no attack trivia (e.g. “what is drop-catching”).
• No computing PCA or t-SNE by hand; no asking for the difference between ridge and lasso.
• Throughput vs. latency may appear on every exam: a small trace or request/response timeline, compute both.

Output

• Instructor: write to the Dropbox directory above. After the exam, publish it with solutions in docs/combined/<YYYY>/ and add the term’s prompt record (see generate-summer-final.md for 2026) if the parameters were changed for that year.
• Students: any directory of your own, e.g. ~/ml-systems-practice/combined/. Copy feamster.sty and the Makefile from the most recent past combined exam, then make all.

Validation

As in generate-exam.md: make all, check that pdfinfo exam.pdf | grep Pages gives 6 (or 7 if the feedback item spills) and that the \prob{} points sum to 75; open the PDF and look for overflowing option text or answer boxes broken across pages.