Assignments & Grading
The opening labs establish inference and prompting baselines. The individual homework then progresses from reproducible compute practice to supervised fine-tuning and preference optimization. The group project asks your team to integrate those ideas on a task you choose.
Grading Breakdown
| Component | Weight | Released | Due |
|---|---|---|---|
| Participation and in-class labs | 10% | ||
| HW0: TACC inference and resource accounting | 5% | Aug 26 | Sep 9 |
| HW1: Supervised fine-tuning and LoRA | 15% | Sep 23 | Oct 7 |
| HW2: Preference data and DPO | 15% | Oct 21 | Nov 4 |
| Concept check (in class, handwritten) | 10% | Nov 11 | |
| Group project proposal | 5% | Sep 23 | |
| Group project checkpoint | 5% | Oct 28 | |
| Group project final presentation | 10% | Nov 18 | |
| Group project revised final report and code | 25% | Dec 7 |
Plus and minus grades are used in the final class grade. Final percentages are not rounded up.
Homework
HW0: TACC inference and resource accounting 5%
Use a provided SLURM workflow to reproduce course-designated model inference on TACC, inspect job outputs and accounting information, and estimate compute use. The emphasis is on reproducibility and responsible use of shared resources.
HW1: Supervised fine-tuning and LoRA 15%
Fine-tune a course-designated open-weight model with LoRA on a provided dataset and compare it with a prompting baseline. The graded core is a correct, controlled experiment and an evidence-based interpretation of the training behavior and evaluation results.
HW2: Preference data and DPO 15%
Use a provided pool of prompts and candidate responses to make and document your own preference judgments, then run a controlled DPO experiment and compare it with a reference checkpoint. The emphasis is on human labeling, controlled evaluation, and interpretation of behavioral changes and possible regressions.
Concept check 10%
A short, handwritten, closed-device assessment covering the core concepts of the first eleven weeks. It exists to confirm the core ideas landed.
Attendance & Participation
Because the class meets only once a week and much of the lab work happens collaboratively in the room, attendance matters. The 10% participation component reflects attendance and engagement in labs. You may miss up to two class meetings without affecting your participation grade. After those two, each additional unexcused absence lowers the final course grade by two percentage points through the participation component.
Absences for illness, religious holidays, university-sponsored activities, or documented emergencies are excused and do not count against your participation grade or your two permitted absences. If you miss a lab, the notebook and instructions will be on Canvas, and the TA can help you catch up in office hours.
Late Work & Makeup Policy
Each student has four late days for the semester, usable in whole-day increments across HW0, HW1, and HW2, with no penalty and no explanation required. Beyond that, late work loses 10% per day. Project milestones are not eligible for late days.
Final Examination
There is no final examination in this course, and none is scheduled during the University final examination period. Final project presentations take place on November 18, 2026. The December 2 class is devoted to feedback-based project revision and refinement, and the revised final project report and code are due at 11:59pm on December 7, 2026.