Final Project
Group Mini-Project β LLM-Based Linguistic Analysis
The final project is a group mini-project in which you run a small, end-to-end LLM-based analysis on a real annotated dataset and report the results. It is introduced in Session 3 (track selection + group formation) and carried out across Sessions 12β15.
Choose a track
Pick one of the provided expert-annotated datasets (easy β hard). Details and licensing are on the Replication Datasets and Mini-Project Starter Tracks pages.
Workflow
- Sample a balanced ~40-item gold subset from the provided pool (
sample_pool). - QC / adjudicate the sampled gold set.
- Iterate your prompt over 2β3 rounds, re-evaluating each round (P/R/F1, confusion matrix).
- Freeze your final predictions to JSON (see below).
- Report in an in-class one-page report.
- Present to the class with instructor Q&A.
You run the model through the Gemini API with temperature=0 + a fixed seed (get a free key). Even so, a hosted LLM is only best-effort reproducible, so once your prompt is final, run it once and save the predictions to a JSON file, committed with your notebook. Your evaluation then runs off that frozen file β so your reported F1 is stable, and anyone (including the grader) can re-run your analysis on exactly the outputs you saw.
Deliverables
- Presentation + Q&A β the graded core.
- In-class one-page report β track/labels Β· gold + what QC changed Β· prompt-iteration table (F1 per round) Β· confusion matrix + error analysis Β· limitations.
- Completed notebook β assembled from the cell library and run end-to-end, with your groupβs gold subset, prompt, frozen predictions JSON, and outputs.
All deliverables are produced and submitted during the course β there is no post-course write-up.
Grading
See the Course Syllabus for the full breakdown (mini-project presentation + Q&A and the in-class report are weighted components; the completed notebook is graded under hands-on).