Session 13: Project Work — Prompt Iteration & Final Evaluation

Day 5 · Project Finalization & Presentations (5-1)

Warning🚧 Being prepared

This page is being finalized ahead of the course (Aug 3–7, 2026) and may be incomplete or change before your session. The syllabus and readings are ready now.

Learning objectives

By the end of this session you will be able to:

  • Change one thing per prompt round, predict its effect, and test it against your gold set.
  • Read a confusion matrix to decide what to change next, rather than guessing.
  • Freeze a final run to a file, and explain why a seed alone is not enough.
  • Report plain and weighted κ, and say which one your label set justifies.

Agenda

  1. Round 1 and 2 — few-shot, then a change of your own. For each: what did you change, and why did you expect it to help?
  2. Read the errors between rounds — the confusion matrix tells you which pair to attack.
  3. The final run — raise N_PER_CLASS, run once, save_predictions.
  4. Everything after that reads the frozen file, not the model.

Budget your quota: the free tier gives about 500 calls a day per key. Iterate small.

Reading

No new reading — Day 5 is project work only (see the Readings page).

Slides & Colab

Mini-project

04_develop.ipynb — the prompt rounds, on the dev half — then 05_test.ipynb for the held-out run, once.

ImportantThe final run is one run, by one person

Runtimes are per-person, and your reported numbers have to come out of a single frozen file. Decide who the driver is before you start. If their daily quota runs out, hand the role — and the runtime — to someone else; the files stay put in your shared Drive folder.