User guide

TA workflow

The Streamlit app is the main interface. It configures a GradingService and shows what comes back, and nothing in the UI grades anything itself.

Select an assignment
  → upload a Canvas ZIP or point at a submissions folder
  → review what discovery found
  → Run Grader (notebooks execute in isolated containers)
  → class overview: mean, distribution, most common issues
  → review queue: only the submissions that need a decision
  → per-student page: rubric scores, evidence, manual override
  → export grades.csv and per-student feedback

Start the app#

streamlit run app.py

The app binds to localhost. Uploads can be up to 600 MB, because the Zillow extract is well over Streamlit's 200 MB default. See .streamlit/config.toml.

The sidebar#

Assignment#

Choose any rubric found in rubrics/. The list is built from every YAML file that parses.

Submissions#

Upload a Canvas export ZIP, or type a local folder path and press Load folder. Both layouts are supported: a folder of per-student folders, and a flat Canvas export (doejane_12345_67890_assignment1.ipynb, including _late_ files).

Assignment data#

Students hand in a notebook and nothing else, so you provide the data. Upload the file(s) here, or use "or use a file already on this machine" to give a path. Uploads are kept in .musa_grader_data/ and survive a restart. Each file in use is listed with its size and can be removed individually.

Pages#

The review queue#

A submission goes to the queue when any of these applies:

• It lost any points at all. This rule is on by default, and it is the strongest one.
• The notebook failed to execute or timed out.
• Discovery found multiple notebooks or none.
• An item's confidence is below the threshold (default 0.80).
• An item has status manual_review. This includes every qualitative item when LLM grading is off.

Each queue entry names the items that lost points and how many, and broken submissions sort above small deductions. Marking a submission reviewed, or overriding it back to full marks, removes it from the queue.

Manual overrides#

An override never replaces the automatic score. Both are kept:

{ "automatic_score": 12, "manual_override": true,
  "final_score": 15, "override_reason": "Alternative valid implementation." }

Re-running the grader regenerates the automatic scores. With Preserve manual overrides on, your decisions are reapplied on top. For extra-credit items (worth 0 points in the rubric), an override adds to the student's total without changing the maximum they are graded out of.

Settings#