CuraCourseAt a glance
AI grading works best when you treat the model like a trained teaching assistant: give it a clear job, explicit standards, and examples of what good and weak work look like. Put shared policy in Grading Context, and put measurable dimensions in Scoring Criterion. These practices match the Info help inside the AI Grading item in the lesson editor.
Prerequisites
- An activity with an AI Grading rubric item. See Add rubric items (manual, AI, and code grading).
1. Separate instructions from criteria
Use Grading Context for global rules: course norms, tone, partial-credit philosophy, and what counts as full credit.
Use Scoring Criterion (and the Criterion Name) for one measurable dimension — accuracy, reasoning, completeness, use of a required method, and so on.
Mixing both in one block makes scores harder to reproduce. Put shared context on the first AI item in the group; later items can add extra context if needed.
2. Write criteria as observable behaviors
Describe what appears in the student’s answer, not what you hope they understood.
- Prefer: names the correct algorithm and explains time complexity
- Avoid: shows good understanding
Observable language reduces drift between runs.
3. Define every point level
Weight sets both the item’s points and how many levels you must describe (0 through that number). Fill every level. The model anchors on these bands; vague middle descriptions produce inconsistent partial credit.
When in doubt, add a short example phrase at each level (what a 0, 1, and 2 actually look like).
Keep Weight modest at first (2–3 points). Changing weight resizes the descriptor slots, so write levels after you settle on the scale.
4. Keep criteria independent
Add a separate AI item per dimension instead of packing unrelated ideas into one descriptor. Avoid double-counting: if “correctness” and “uses proper terminology” overlap heavily, merge them or narrow each so one mistake is not penalized twice.
5. Calibrate with test grading
Before you assign the lesson, submit sample responses that represent a clear pass, a borderline case, and a fail. Adjust descriptors until scores match your intent. Re-test after you change Grading Context, keywords, or point totals.
6. Tell the model how to handle ambiguity
In Grading Context, state whether to:
- Reward reasonable paraphrasing
- Deduct for missing optional details
- Enforce strict spelling, units, or notation
Explicit policies prevent harsh or overly lenient outliers.
7. Request actionable feedback
Ask for brief student-facing feedback tied to the criteria (what was strong, what to fix). That helps learners, and it helps you see when the rubric itself needs work.
Instructors also receive private Instructor Info notes in grading results. Use those to audit the model; do not rely on Lightning Grader to edit the AI score row — adjust with score modifiers or general feedback if needed.
Keywords
Keywords are optional hints, not a checklist the student must hit verbatim. Use them for course-specific terms you want emphasized. Do not expect all-or-nothing keyword matching.
Start simple
Two or three criteria with clear levels is enough. Add complexity only when test grading shows a consistent gap.
On a Code Editor, AI grading reads the submitted files as text. Use it for design, style, or explanation — not as a substitute for Code Grading tests that check actual output. See Best practices for code grading.