Model Evaluation
What I built
Reference answers, question difficulty tiers, test contamination checks, model and judge adapters, grading rubrics, agreement checks, sampled regression comparisons and saved checkpoints.
Result
Built reproducible grading, calibration and regression analysis. The model judge agreed with 79.3% of 150 constructed reference labels.
Evaluation setup and selected results
On 150 constructed reference labels, the model judge agreed with the reference grade in 79.3% of cases. Agreement adjusted for chance was 0.628. The tooling supports grading-rule calibration, disagreement analysis, repeated-sample comparisons and saved regression checks. This measures agreement with the constructed labels; it does not establish factual answer accuracy.
Compare the evaluation evidence →
About the evidence
Based on source review and saved project artifacts. Recorded measurements apply to the stated model, dataset and configuration; they are not claims of a current production deployment.