eval-metrics

What problem does it solve?

Automates the computation of evaluation metrics for ordinal predictions (e.g., Spearman ρ, Kendall τ), providing exact and adjacent match percentages, mean scores per level, and a confusion matrix. Outputs: evaluation/{year}/metrics.json. Pure statistical computation — no LLM calls.

Core Features & Use Cases

• Compute rank-based correlations for ordinal grading data.
• Generate per-level average predicted scores and a confusion matrix for error analysis.
• Support optional level divisions and resume-safe operation across grading years.

Quick Start

Run the evaluation workflow to generate metrics from ground truth and final scores for a specific subject/year.