Calibrated Human-in-the-Loop Short-Answer Grading
A fine-tuned language model grades student responses and emits a temperature-scaled confidence score. High-confidence predictions are auto-graded; low-confidence ones are flagged for human review. Attribution highlights the answer tokens that most influenced the grade.
Examples:
Loading model and running inference. This may take a minute on first request.
-
Predicted Grade
-
Confidence
Token Attribution
Gradient x Input: tokens most influential to this grade
LowHigh attribution
Model Feedback