Evaluation vs Baseline

Decide whether your fine-tuned model ships — beat the baselines, measure per-class accuracy and schema validity, test significance with bootstrap intervals, check forgetting, count cost and latency.